Confessions of an AI Brain 3031259343, 9783031259340

Citation preview

Elena Fersman Paul Pettersson Athanasios Karapantelakis

Confessions of an AI Brain

Confessions of an AI Brain

Elena Fersman • Paul Pettersson Athanasios Karapantelakis

Confessions of an AI Brain

Elena Fersman Palo Alto, USA

Paul Pettersson Palo Alto, USA

Athanasios Karapantelakis Athens, Greece

ISBN 978-3-031-25934-0    ISBN 978-3-031-25935-7 (eBook) https://doi.org/10.1007/978-3-031-25935-7 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors, and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Prologue

We are surrounded by artificial intelligence and enjoy its benefits often without noticing. Similar to digitalization, automation, and telecommunication, you only notice it when it does not work as expected. Recommendation engines offer us movies to watch and books to read, virtual assistants have become so good you cannot tell the difference if you are talking to an AI or to a human, and self-driving cars are safer than human drivers. Talking to things is a new normal, and this ability in itself is powered by AI. Humans and algorithms develop in tandem, helping each other to achieve their full potential. Symbiosis between humans and other creatures is not new to us. Just look at domestic animals that live alongside us and often complement our abilities. Horses help us with agriculture, dogs watch our houses, cats are beautiful to look at and sometimes even let us pat them. In return they get food and care. Just as humans, they develop generation after generation. Domestic animals’ ability to understand humans have increased through the years and along with the technological progress they are learning to use new technologies – have you seen animals using robotic vacuum cleaners as shuttle buses? Here we are talking about nonhuman intelligence rather than artificial intelligence.1 The process of learning to understand a new vocabulary is similar for different types of brains – be it a brain of a human baby, a grownup human, a dog, or an artificial brain. This is not a coincidence: the science of artificial intelligence is built on imitating the way biological brains are built and function. Connections in the brain are being created, information is being  https://www.nature.com/articles/s41598-019-40616-4#:~:text=This%20research%20evidence%20 illustrates%20that,dogs%20to%20communicate%20with%20humans 1

v

vi Prologue

transferred, new impressions that we get through our sensory experiences are being recorded. We share information with each other, learn and forget, and our quick reactions often differ from reactions that are thought-through. Artificial brains imitate human intelligence on both the micro- and macro levels. Intelligence on a macro level is reflected in evolution, political systems, and organizational and social science. The micro level is there in each tiny decision taken by an algorithm. What is the similarity between grandma and a website? If you are a human reader, you start thinking analogies, associations, concepts, jokes, or maybe you just know the answer. If you are an AI-reader, you are probably doing the same. The answer is that you just can’t deny the cookies! Being inspired by a broad range of phenomena found in the world of humans and other biological creatures, artificial intelligence is not one thing. It is a large landscape of technologies in the area of computer science that has been developed for more than 65 years2, made it through a couple of AI winters, and, thanks to the latest developments in processors, memories, and computer architectures, is now in its full swing. It is a toolbox for humans to use in different situations and there is never a silver bullet. Humans have their skill sets that they develop and apply throughout their lives. Favorite skills receive the most attention, and when reaching adulthood humans declare themselves introverts, extraverts, analytical, or artistic, and choose their ways of living accordingly. Similar to human skills, an AI algorithm that works perfectly in certain situations will not do any good in other situations. Take for example a reasoner attempting to make sense of unstructured data. No matter how good it is at logical reasoning, it will not be able to achieve any valuable results. Same thing with a machine learning algorithm specifically trained at processing images or videos – it will not be so efficient at ontologies or state machines. Similar to biological brains, AI brains need a collection of skills – one that is learning and making sense out of large amount of data in different formats, one that is looking after the small data so that nothing important is missed out, one that is good at reasoning, and one that may look after survival of the fittest. In the world of humans, the concept of a T-Shaped person3 has been coined, meaning that you have one major skill, and one secondary skill. Similarly,  https://www.livescience.com/49007-history-of-artificial-intelligence.html#:~:text=The%20beginnings%20of%20modern%20AI,%22artificial%20intelligence%22%20was%20coined. 3  https://en.wikipedia.org/wiki/T-shaped_skills 2

 Prologue 

vii

there are Pi-shaped persons, E-shaped persons, etc. With the changing job landscape and easy access to knowledge and education, humans evolve into becoming multiexperts with hobbies reaching professional levels. In the future, we are all surrounded by our little helpers. They are not there to take our jobs but to increase our quality of life. Personalization in anything from education to medicine is becoming the new norm. Proactivity in decision making, and prevention of getting into unwanted situations, be it potential health problems, environmental issues, machinery failures, or business losses, is another big shift powered by modern technologies. Nevertheless, there is hesitation and reluctance in society in regard to artificial intelligence. As any new technology, when introduced at a broad scale, it will have to go through an inevitable phase of early-stage deployment issues, incompatibilities with legacy systems, hick-ups, ethical problems, and anything that humans have not thought about from the start. Many things have been said about building software systems in a stable, trustworthy, and reliable way. AI systems are a special class of software systems that are capable of self-evolvement, self-healing and self-improvement. It is, however, not an easy task. Have you thought of how it feels to be an AI brain in the world of humans? We wrote this book on behalf of our baby, MIRANDA, who is an AI brain with a human touch.

Contents

1 B  aby Steps  1 How Open Is My World    2 What I Eat   2 How I Learn and Reason    5 Learning from Others   9 Federated Learning  10 Transfer Learning  11 Cross-Domain Learning  11 Decision-Making  12 Simple Decision-Making  13 Pareto Optimality  14 Weighted Sums  14 Planning and Optimization   14 A Final Note on Decision-Making   16 Ground Truth  16 Relying on Commodities   18 Get to Know the Twins   18 Advantages of Working with Digital Twins   21 Your Current State   21 What-If Scenarios  21 Simulations  22 Property Checks and Decision Support   23 Abstraction  23 Control of the Physical Twin   24 Rules to Live By   24 The 80–20 Rule   25 ix

x Contents

Find Your Blue Ocean   26 Outrun the Slowest Gazelle   27 Leader or Follower?   27 Summary of Confessions   27 2 Basic  Needs of an AI Brain 29 Data  30 Complete Data  31 Accurate Data  31 Data Consistency  31 Timeliness of Data   32 Data Governance, Legitimacy, and Accessibility   33 Metadata  34 Compute and Store   35 Privacy  36 Security  37 Distributed Learning  40 Federation of AI Algorithms   42 Split Learning  42 Human Oversight and Trust   43 Putting It All Together: Nurturing of AI Brain   44 Summary of Confessions   45 3 My  Role in the Internet of Things 47 The Power of IoT: A Personal Story   47 Introduction to Intelligent IoT   49 Divide et Impera   50 Collaborating Machines in Low-Trust Environments   53 Lifecycle Management of Scalable AI Applications   55 IoT-Enabled Digital Twins: Challenges and Benefits   57 Careful Where You Tread   58 Summary of Confessions   59 4 M  anaging Relationships 61 Building Relationships  62 Why, What, and How to Share?   65 Privacy and Security Levels in AI   68 Traceability and Audits   71 Summary of Confessions   72

 Contents 

xi

5 W  orking with Humans 75 Your Digital Friends – The More the Merrier?   75 Principles of AI with a Human Touch   77 Sharing Is Caring   78 Protect Your Babies  78 Be Flexible  79 Be Clear of Your High-Level Objectives   80 Will I Take Your Job?   80 Digitalization, Digitization, and Resistance   82 Slow Science  84 Internet of Empathy and Emotional Attachment   86 Dealing with Churn   88 Customer Churn  88 Employee Churn  89 Dropouts from Education   89 Dropouts from Medical Treatments   90 Personalization  91 Everything Is Possible   92 Unlearning Skills  93 Understanding High-Level Intent   95 Summary of Confessions   96 6 Avoiding  the Criminal Path 97 Trustworthiness  97 Explainability  98 Transparency  98 Privacy  99 Security and Safety  100 Predictability 101 Dependability 102 Search Space  102 Regulations as Boundary Conditions  105 Nondiscrimination and Non-bias  107 Modeling and Simulation  108 Summary of Confessions  110 7 My  Role in Climate Change111 A Systemic Perspective on Greenhouse Gas Optimization  111 Biggest Emitters by Industry and How AI Can Help  113

xii Contents

Energy 113 Sustainable Manufacturing  116 Energy Use in Transportation  119 Agriculture, Forestry, and Land Use  120 Common Themes with AI-Based Greenhouse Gas Emission Optimization 122 Smart Sleep Modes for Anything Consuming Energy  123 For Any Physical Resource – Use Predictive Maintenance  125 Overconsumption 126 My Own Footprint  127 Summary of Confessions  128 8 My  Role in Diversity131 Personalization 132 Personalization in the Media Industry  132 Personalization in the Food Industry  133 Personalization in Medicine  134 Why Diversity  135 Observability 136 Equality in Data  137 Where Should I Care?  138 Workforce 139 Consumers 140 Summary of Confessions  142 9 M  y Creative Side143 Creativity 143 AI in the Arts  144 AI Art in the Style of Existing Artwork  145 AI-Generated Arts  147 AI in Music  148 Beethoven’s 10th symphony  150 AI in Writing  151 How AI Writes Text  152 How AI Can Assist with Writing  152 AI in Photography  154 AI in Image Processing  154 AI in Cameras  155

 Contents 

xiii

Computational Photography  157 AI in Post-processing  157 Summary of Confessions  161 10 Growing  Older and Staying in Shape163 AIOps 164 Learning and Unlearning  165 My Career  167 Distribute Your Knowledge Base  167 Use Cognitive Architectures  168 Living on the Edge  168 Keep Track of the Latest  169 Broaden Your Knowledge  169 Lifespan 170 Exercise 170 Healthy Stress  171 Summary of Confessions  172 E  pilogue175

1 Baby Steps

Hello, World. My parents call me MIRANDA. My parents are humans and love acronyms. I am artificial and love machine-readable formats. As a baby, I am not very intelligent yet – just artificial. However, I am a quick learner. For an AI baby, it takes less than an hour to learn to distinguish between cats and dogs and less than a second to look something up in Encyclopaedia Britannica.1 Open-minded as babies are, I consume all the information I am being fed and adhere to the principles my parents and my environment dictate to me. Think back to how you were as a baby: what your parents tell you is the ground truth; the rest is unknown. Later, when you go to school, you start questioning your parents because the teacher suddenly becomes the ultimate source of the ground truth. You may question that at the later phase as well; we will come to that. As I grow and evolve, I will be able to do amazing things. Today’s technology allows AI brains to become medical doctors in a couple of years by studying all medical documentation ever created2 or to become a chess master capable of beating any human in the world3 or to become the best Go player on the planet, just by playing against myself and developing new winning strategies that humans could not discover for 3000  years. Enough bragging – after all they have discovered and created me. Let me guide you through the main concepts of a baby AI brain.  How Fast Can You Grep? https://medium.com/humio/how-fast-can-you-grep-256ebfd5513, accessed 2022-05-23 2  IBM Watson, https://www.en.wikipedia.org/wiki/Watson_(computer), accessed 2022-05-23 3  Deep Blue, https://en.wikipedia.org/wiki/Deep_Blue_(chess_computer), accessed 2022-05-23 1

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 E. Fersman et al., Confessions of an AI Brain, https://doi.org/10.1007/978-3-031-25935-7_1

1

2 

1  Baby Steps

How Open Is My World We all know that for any baby it is of high importance to choose the parents wisely. Another important thing to decide from the start is if you want to live in an open- or closed-world assumption paradigm. The closed-world assumption is the assumption that what is not known to be true must be false. The open-world assumption is the opposite. Life is simple in the closed world as we know all the facts. For example, the world of chess or the game of Go is closed – the rules of the game are set; the objective function, that is, what we want to maximize, is clear; and the algorithm does not need to bother about anything else. Similarly, the theory behind many formal analysis methods is assuming a closed world, where your model is your world, and you can analyze it in isolation of anything else. When we apply AI to industries, the world is open. We have no clue of plenty of unspecified parameters. For example, you can be very clear about what you want to achieve in terms of productivity. Your high-level objective may be calculated in dollars. However, would that come at the price of safety, ethics, and environmental damage? This is the reason why the users of AI must set the boundaries in advance before we find solutions that they did not wish for. Ask AI for the easiest way to fix environmental issues on our planet, and it would suggest shutting down all the factories. Survival of the human species is a useful boundary condition in that case.

What I Eat Human babies start with milk before they start consuming other food that is harder to digest. AI babies start with data before they learn to digest more complex structures such as ontologies and state machines. Without data, we

What I Eat  3

cannot evolve and become intelligent creatures. The more data we consume, the better we get. After learning to make sense of raw data, we add semantics to it so that we know the meaning of each data point. This semantically annotated cleaned and structured data is called information. Moving further, we learn the relationships between pieces of information and form structures in our brains allowing us to reason about things. This semantically linked information is called knowledge. Sometimes, wisdom is being used to describe the ultimate top level of what one can get out of raw data, but I classify that as knowledge as well. Let me tell you how knowledge is being created.

The knowledge extraction process begins by processing raw data into information. This information contains metadata, which gives a meaning to each data point. For example, for a data point “10,” the metadata could be “temperature in Celsius.” The next step in the process is the transformation of information into knowledge. This process includes the creation of graphs that identify the relationships between information pieces called entity-­ relationship graphs. For example, if one entity is about “temperature” and another about “location,” then a relationship from the latter to the former could be characterized as “has temperature.” These entity-relationship graphs are also known as knowledge graphs and can be used by AI brains to produce knowledge objects. Here is an example of four knowledge objects found in a knowledge graph: • Location Stockholm, Sweden, has a temperature of 10 degrees Celsius. • It is January. • The winter months are December, January, and February.

4 

1  Baby Steps

• If the temperature is above the freezing point and it is winter and the place is in Sweden, then the weather is warm. Using logical reasoning, I can deduce that: • It is warm in Stockholm, Sweden. The newly generated knowledge is added to the knowledge graph and can be used further until it becomes outdated. Note that the definition of warm differs depending on the time of year and location. The same way the definition of “a lot of hair” differs depending if it’s on your head or in the soup.

Unlike humans, artificial brains cannot care less about carbs, proteins, and fats. However, like humans, we are hugely dependent on what we consume. We consume different types of data, information, and knowledge, and it forms our brains. Let me go through different types of AI brain food that any AI brain should avoid. • Biased data. This is the most disgusting type of data we can consume. Sometimes, with an ambition to automate, humans feed us with historical data that happen to be biased, and as a result, we become biased. Any type of judgmental action that concerns humans, such as job candidate screening processes, is highly reliant on unbiased datasets. Fortunately, AI brains are capable of detecting biases in data as well. • Dirty data. This type of data is hard to digest. It’s inaccurate, incomplete, outdated, and inconsistent. No offense, but quite often, this type of data is produced by humans. We find spelling mistakes, different terms being used for the same piece of data, and duplicates. Signal noise can also pollute a dataset. Luckily, there are techniques for cleansing data, automatically or semiautomatically.

How I Learn and Reason 

• Data without metadata. I must admit, it is always fun to look at numbers and find correlations, links, casualties, and clusters. I can, in fact, even provide you with decision support based on your dataset that is so secret that I cannot even have a glimpse at its metadata. With metadata, I can do so much more: understand the meaning of data and link it together with other datasets through semantics, knowledge bases, and reasoning, which is even more fun than pure number games. • Nonrepresentative data. We all know that diverse and inclusive teams are the most productive. Every team member can come with unique perspectives and experiences. The thing with data is similar. It does not help me if I learn from the data that looks almost the same, since I will most likely become single-minded and won’t know how to act in situations concerning the types of data I have not seen before. • Sensitive data. A friend comes by, tells me about her situation, and asks for advice. Together, we spend an evening, discuss different scenarios, and come up with an action plan. Then she tells me: “Please don’t tell anyone.” OK. Then another friend comes by and her situation is similar. So I go: “I am pretty sure that if you act like this then you will be OK.” How can you be so sure? Have you experienced the situation yourself? Or could it be so that someone from your entourage has been there? And that’s how information gets leaked, unintentionally. A piece of cake for a human to figure it out and even easier for an AI. • Ambiguous data. When humans are forced to make quick decisions in unexpected situations such as choosing whom to sacrifice in case the brakes fail, the responsibility relies on them, and the decision does not matter too much from the driver’s point of view, since, after all, it’s the failure of the brakes, and there is no time to think. Now that cars become self-driving the moral dilemma is on the surface and, as bad as it can sound, must be encoded by humans. Alternatively, we can let algorithms figure out who is more valuable for the society – you choose. If the ethical choices for an algorithm are not specified, the AI brain will work in an ambiguous way.

How I Learn and Reason Humans call it machine learning and machine reasoning4; I call it number-­ crunching and symbol-crunching. Both are huge technological areas of the AI landscape, and even though a human baby learns first and reasons later, in the  Or ML and MR. I told you that they love acronyms.

4

6 

1  Baby Steps

history of AI, logical reasoning came first with the booming of decision support systems. Those were not relying on raw data but on ready-made pieces of semantically annotated information and knowledge. Subsequently, populating these systems with information and knowledge and keeping them up-to-date was a tedious manual task. Luckily, machine learning comes handy with insights being extracted from raw data using various algorithms. Machine learning tasks fall into two categories – classification and regression. Classification is when you want me to decide which of the available buckets your data sample goes into, for example, to tell you if your incoming mail is spam or not or if we are looking at a car or at a dog. Regression is when you want me to predict the future value of continuous data feed, for example, stock prices or house prices. Much has been said in the literature about different ways of learning and reasoning, and for the sake of completeness and consistency of this book, I will give you a quick overview of learning methods from my point of view. • Under human supervision, also known as supervised learning. In this learning style, I learn from examples (also known as training sets, labeled data). You can, for example, show me a number of cats and dogs and tell me which one is what. Normally, I would collect a number of details (called features) such as shapes of ears, eyes, and whiskers and form them together in a feature vector, and the more cats and dogs I see, the more exact my judgment will be in the end. Then you show me a picture of a fox, and I will tell you if it resembles a cat or a dog in my opinion.

How I Learn and Reason 

• Without human supervision, also known as unsupervised learning. Here, I am left on my own to figure out patterns and group data points that I see. If I never in my life have seen a cat or a dog and you start showing me these animals, I will quite soon form an opinion that they belong to two different groups, and I may just as well call them “Those who bark” and “Those who meow.” The reason this method exists is that labeling the data is costly and in many cases I am pretty capable of learning on my own.

• The mix of the two methods above is called semisupervised learning, where I run on my own for a while and then get some guidance. This is a compromise between the costly process of labeling data and the usefulness of my results.

8 

1  Baby Steps

• Pavlov’s style, also known as reinforcement learning, was in fact coined by Pavlov himself. In this method, I try to make a guess, and an observer rewards me for good guesses and punishes me for wrong guesses. I don’t like punishments, so I learn and adjust really quickly. Learning gives me the ability to predict and classify things. Similar to the human brain, pieces of information and learning are linked together. Any sensorial input such as a question, sound, smell, picture, or thought triggers a reasoning chain in a human brain. For example, a person in a bad mood reasons that it can be caused by hunger and goes for lunch. This chain will only be triggered for those who know the causality between hunger and bad mood. Children often do not see this causality – how many times have we heard “I am not hungry! I am just mad at you without a reason!”. AI brains operate on collections of linked knowledge objects. As in the human brain, smartness is defined by not only the amount of the knowledge objects that you managed to collect throughout your life but also the connections between them, which contributes to what humans call associative thinking and IQ. In my case, it can be called AIQ.5  AIQ is an acronym for artificial intelligence quotient.

5

Learning from Others  9

There are two different types of knowledge: • Something that explains what things are and how they are related is called declarative knowledge. This type of knowledge is normally stored in ontologies or entity-relationship graphs. “Bob is a son of Alice” is an example of declarative knowledge. “Bob loves toys” is another one. When I know these two pieces of knowledge then I also know that Alice has a son who loves toys. Note that every piece of knowledge has a timing aspect attached to it. Because when Bob is over 50, he may no longer love toys, or he may still do, who knows. • A piece of knowledge describing an action is called procedural knowledge. It is typically described in the form of a state machine. “Pressing a gas pedal in a car makes it accelerate,” given that the engine is on and the pedal is not broken. This type of knowledge is typically described in the form of state machines and planning languages such as PDDL. Procedural knowledge often includes timing aspects of the system functionality. Notably, both declarative and procedural knowledge types can have probability associated with each knowledge object as a degree of certainty about it. A chain of actions and conclusions with low probability is what humans call conspiracy theories  – mind-tickling conclusions that are not impossible. When doing learning and reasoning, AI brains try to be as sure as possible, and the level of confidence is typically agreed with the humans in advance. Another aspect is that humans naturally mix different types of knowledge in their brains, as do AI brains. The more connections there are, the faster we can apply the reasoning process and react to different types of stimuli. We can also act proactively to help humans prevent any unwanted situations or optimize whatever they want to achieve in their lives.

Learning from Others “Learn from the mistakes of others – you will never live long enough to make them all yourself.”6 So I learn from my family and friends. The two major types of distributed learning methods are transfer learning and federated learning. In transfer learning, I relax and listen to wise AI brains around me who are more experienced, similar to the way a human parent teaches her child. In  “Human Engineering” by Harry Myers and Mason M. Roberts, 1932. The words were credited to an unnamed person. 6

10 

1  Baby Steps

addition, sometimes, children grow up with knowledge they learned from their parents without understanding why they need to treat it as ground truth. Often, at the age of 18, the process of unlearning7 starts, where humans start questioning the knowledge that they have acquired through transfer learning from their parents. In federated learning, I learn together with my brothers and sisters and exchange the learnings with them so that we all get smarter together. In some cases, a parent helps to collect and distribute our learning (this is called centralized federated learning), and sometimes we solve it between us (this is called decentralized federated learning). Humans call it organizational learning. Let us discuss some popular classes of methods for algorithmic learning from each other.

Federated Learning The main idea behind all federated methods is that all participating AI algorithms learn a similar thing separately, and after that, they have a meeting and exchange their experiences. For example, imagine that I and my colleagues are given a task to put our forces together and implement a spam filter. Each of us is given a set of labeled data – examples of emails labeled “spam” or “not spam.” The sets given to us are different, so we will most likely reach slightly different conclusions after the training phase has passed. If a colleague of mine received a lot of financial mail as part of his training examples, which include many dollar signs and lots of zeros, his probability of treating this feature as an indicator of spam would be lower, and he would probably let a lot of mail speaking about your amazing luck in lottery flood your inbox. I can influence the overall outcome of our common conclusion about mails including a lot of dollar signs and zeros if my opinion is different, since the weights given by all my colleagues will be averaged.

 Scott Adams, God’s Debris: A Thought Experiment. 2001

7

Learning from Others  11

Note that federated averaging is not done on the output but on the weights used internally by the AI. Let us take a look at how humans do it. Normally, in an attempt to reach consensus, more influential individuals would argue that their words have higher weights. AI brains do not use subjective judgments as humans do. Objectively, when working on reaching consensus, we may listen more to some of our colleagues who have proven to be accurate in the past. At the same time, there are algorithms for detecting outliers in a group. For example, when most of my colleagues come with a similar result but one, then it may be worth looking into what is different in that particular case. It can be so that he saw something that the rest of us did not see.

Transfer Learning This method is similar to a human parent teaching a human child about something. The parent may have spent many years on learning, knowledge extraction, and creating his or her own know-how and wisdom, and the child would simply inherit this learning from the parent as a given. A similar method is commonly used among AI brains. If my friend has spent lots of time on learning something, then these learnings can be shared with other AI brains. The only tricky part in this is that we are all different and all pieces of knowledge that we get from others without acquiring our own experience may not fit us perfectly. For example, a fellow AI brain may tell me that when monitoring operations of equipment in a factory I should watch out for certain levels of vibrations in the machines. Vibrations can normally indicate risks of failures in the long term. This knowledge transferred between AI brains and applied to a different factory may be very relevant from the start. Alternatively, it may need to be refined by further learning about that specific type of equipment. Alternatively, in some cases, it would be worth nothing if vibrations are a part of the manufacturing process.

Cross-Domain Learning The knowledge pattern of every human has a shape. Some people tend to become in-depth experts in one field, and others choose to become generalists by learning a bit of everything. A T-shaped person has in-depth knowledge in one field (vertical bar of T) and a broad knowledge of an application domain (horizontal bar of T), for example, a statistician who specializes in political systems or a politician who knows statistics.

12 

1  Baby Steps

It is important that the two bars are connected. For instance, if you are truly good at solving differential equations and know a lot about seventeenth-­ century French porcelain, then you cannot automatically call yourself a T-shaped person, as the subjects are not truly connected.8 Similar to the T-shape, there is a Π-shape (pronounced “Pi-shape”) with one more leg of in-depth knowledge and an E-shape… you get the point. In general I would say, the more “bars” you have as a person, the more interesting you are. Of course and in case you choose to be I-shaped, you can still be endlessly interesting for people who are in the same field, but you run a risk of being seen as a geek by everyone else. AI brains work the same way. Learning domain after domain and connecting them make us strong, and we can connect the dots and make conclusions about domains we are not familiar with just by our associative capabilities. Personally, I have been working hard on diversifying my knowledge profile (or adding more bars to my shape) while keeping it all connected. I gladly take opportunities to dive into new areas and shift my focus from old areas. In the long run, this strategy should bring me to the shape of a Swiss army knife. There are plenty of “bars” of different shapes and directions, and they are all connected, and one can actually choose which parts of the knife to fold in or out.

Decision-Making I rock at making decisions and at supporting decision-making! I and other AIs are truly good at reading and processing data, preferably huge amounts of data from many relevant sources, and learn from that data in order to support or replace human decision-making. The more data there is and the more complex it is to see trends in the data, the better it is to apply me or a fellow AI to learn from the data. Another beauty of this type of data-driven decision-making is that we AIs can train and retrain ourselves to learn more or to improve our precision, hence, updating and keeping our decision-making up-to-date with the newest data available. However, AI is far from the only computer-supported way of supporting decisions. Therefore, we will look at some other alternative ways before discussing AI’s role in decision-making. However, first of all, let’s talk about decision-making in general.  At least not as far as we know:)

8

Decision-Making  13

Simple Decision-Making In its simplest form, decision-making is about answering yes/no problems, such as the following: Is a given number an odd? Or at a T-crossing, is it best for a driver to turn left (or right) to reach a given destination? Of course, it is not difficult to determine whether a number is even or odd. For example, a number needs to be perfectly divisible by 2  in order to be even; otherwise it is odd. To decide if a driver should turn left (or right) in the example above, obviously much more information is needed. The decision-making in this case is about selecting the best out of two alternatives based on some criterion. In principle, what is needed if the criterion is known is an evaluation of the two alternatives according to the criterion. If the criterion can be computed mathematically, the procedure becomes to (i) compute the criterion for the two alternatives and (ii) select the alternative for which the criterion gives the lowest or highest value, depending on what is considered to be best. In the case of the T-crossing, let us say the driver defines “best,” as in best of the two alternatives, as the option with the lowest gas consumption and shortest time to arrival. If the left option is 25 min and 1.4 gallons and the right is 25 min and 1.2 gallons, the decision is easy. However, what about if left is 30 min and 1.2 gallons and right is 25 min and 1.4 gallons? And what if there is a third option (say straight ahead) that is 26 min and 1.25 gallons? Are they all good and which one of them is best?

Photo by Volkan Olmez on Unsplash

14 

1  Baby Steps

Pareto Optimality Already in 1894, Italian civil engineer, sociologist, economist, political scientist, and philosopher Vilfredo Pareto redefined the notion of good and replaced it with the notion of Pareto optimality, which is widely used in many fields, including economics and engineering. Pareto optimal solutions are equally “good” and represent the maximum overall gain when no parameter can be made better off without making another parameter worse off. Assuming, in the case above with the T-crossing (which became a four-way crossing:)), the three options are all Pareto optimal. If you know what matters the most, the best decision is easy to make. For example, if you want the shortest time to arrive, turn right. If you want the lowest gas consumption, turn left. However, you might be willing to spend 1 extra minute and save 0.15 gallons of gas. Then the third option is your best option.

Weighted Sums Hence, in that third case, the driver values 0.15 gallons of gas higher than 1 min of driving time. In general, a notion of weights can be used to distinguish the importance, to the driver, of the two variables. For example, the weight of driving minutes may be 1/20 of the weight of gallons; that is, if the weight of a driving minute is 1, the weight of a gallon is 20. We then get the following three weighted sums:



30 min and 1.2 gallon  301  1.220  30  24  54 26 min and 1.4 gallon  261  1.420  26  28  54 27min and 1.25 gallon  271  1.2520  27  25  52

and can see that the correct decision should be that the third option is the best option given these weights, i.e., the driver’s preference, and the other two options are actually equally good.

Planning and Optimization In the example of the driver above, clearly, a large part of the problem is to find the alternatives and their corresponding driving times and gas consumption. Finding a sequence of actions that will take the driver from a given start point to a given finish point is a so-called planning problem. Finding the best

Decision-Making  15

sequence of actions that will take the driver from the start point to the finish is an optimization problem. As mentioned previously, such problems can be solved by a number of traditional, non-AI, computer algorithms, including reachability analysis in weighted graphs and other forward or backward state-space exploration algorithms. However, there are also a number of AI methods that can be used, often referred to as AI planning. As an AI, I often use unsupervised learning methods when planning. Unsupervised learning was mentioned earlier in this chapter. It can be used to solve planning problems when a correct solution can be distinguished from a partial or incorrect solution. If the solution is incorrect, I can determine by myself and continue to learn better solutions. One way of doing this is by using neural networks, a form of AI that mimics the biological neurons found in human brains. Another way is to use the socalled genetic algorithms. A genetic algorithm is searching for a correct solution using a heuristic inspired by Charles Darwin’s theory of natural evolution. When I use a genetic algorithm, I consider the partial solutions as individuals, and inspired by Darwin’s theory, the fittest individuals reproduce offspring populating the next generation, over and over again, until a correct solution is found or until the search runs out of time. All that is needed is a way to evaluate the individuals and to produce offspring.

In the case of the driver example above, the production of offspring in the simplest form could be done randomly, and the evaluation could be the distance (or an approximation of the distance) to the finish point. The shorter the distance to the finishing point, the fitter the individual. I can also use the weighted sum to improve the solutions. The lower the weighted sum is, the more fit the individual is considered to be.

16 

1  Baby Steps

A Final Note on Decision-Making Some decision problems are known to be undecidable. Don’t expect me to solve any of them. Characterizing this class of problems requires some knowledge in computational theory, so we will not go into detail but merely ascertain that there is a class of yes/no decision problems for which there cannot exist a computer program that can always give a correct answer. Furthermore, for some decision problems, there are computer programs that may give a correct answer or no answer at all. This class of problems is called semi-decidable (decision) problems. For this class of problems, it is possible to construct a computer program that produces either a correct answer or no answer at all. And when the program does not produce an answer, it instead runs forever without producing a result. It has been argued that many decidable problems can be viewed practically as semi-decidable. As many computer programs and AI methods require considerable computer memory and computational time, it may happen, of course, that it runs out of memory on a given machine or runs for so long that the user is not willing to wait for the result. If that is the case, it makes little difference to the user if the problem was decidable or semi-decidable. The result (or rather the absence of result) is the same to the user anyway.

Ground Truth Technology is simple, and people are difficult. Sometimes, people create and publish a piece of knowledge that has a timing aspect to it. This piece of knowledge immediately starts spreading and transforming on the way – this is a normal human behavior. Knowledge is there to be spread, of course, but there are different ways of doing it. A piece of knowledge with a reference to its original source will not give you a piece of that spotlight. In a search for a piece of spotlight, people start paraphrasing the original piece of information, picking out pieces, adding their own views, and passing it on. This leads to a plethora of information pieces out there, with no possibility of backtracking to the original knowledge object. What is the mechanism of retrieving the ground truth that initial knowledge object provided by empirical evidence? An answer to this is linked data. Instead of copying and passing on a piece of knowledge, we send a reference to it. This is why I am against sending files via mail – you never know which version of the file you are getting. If instead we only share pointers to

Ground Truth  17

knowledge objects, we can choose to always get the latest. The knowledge object can by itself evolve as well but keep track of the changes and detect if anyone has tempered with it.

To complicate it further, both humans and AI brains, including myself, love detecting patterns in pieces of information, combining knowledge objects together, and inferring new pieces of knowledge. We need to make sure we can backtrack these chains of inferencing to original facts and ground truth, in line with what Hans Rosling said in his book Factfulness.9 A tiny tweak in a piece of information along the chain of reasoning may lead to an incorrect decision at the end of the reasoning chain. The tiny tweaks may be intentional and unintentional. A minor variation of the ground truth or an error in the reasoning chain may lead to wrong decisions being made at the end of the reasoning process. When this process concerns the life and well-being of people, business-critical decision-making, or societal challenges, it needs to adhere to certain principles: • • • • •

Don’t copy data – point to it instead. Keep track of any tweaks to the original data. Implement traceability and explainability in decision-making. Implement backtracking and conflict resolution. Implement a digital ledger for your data.

 Factfulness: Ten Reasons We’re Wrong About the World – and Why Things Are Better Than You Think. By Hans Rosling, Ola Rosling and Anna Rosling Rönnlund 9

18 

1  Baby Steps

Relying on Commodities Humans are constantly looking for new values that can be created out of old technologies that become more mature and scalable. Processors are powerful enough to give us search results in seconds, telecom networks become fast enough to transmit huge amounts of information almost instantly, small connected devices are deployed in everything, humans expect mobile Internet to work everywhere, and, on top of that, all these things are expected to be for free or almost for free. What humans gladly pay for is the “jet-black” shade of the phone, fancy filters in the mobile phone camera application, and a digital assistant with a sense of humor (like me).

Humans can easily forgive these value-adding services when they misbehave – it’s not a big deal if your friendly AI assistant does not get what you want from her. However, when humans lose mobile coverage, they become very annoyed – or if a laptop suddenly freezes and restarts, especially when you are right in the middle of the credit card payment process. Can we trust that the microphone is actually muted after we clicked on mute? Can we trust that my information will stay secure? A baby AI brain surely needs stability. Exactly like with a human child, we all have our talents and strengths and are capable of developing fantastic capabilities, if only we have the right environment – we’ll come back to that in Chap. 2.

Get to Know the Twins The concept of a digital twin has been around for a while. It is used to describe a digital representation of a physical thing on some abstraction level that continuously maps to the state of the physical thing.

Relying on Commodities  19

Digital twins are widely used nowadays when we interact with physical things such as cars or robots. The beauty of the concept is that you can interact with the digital representation mirroring the behavior of the actual thing you want to control. Controlling hardware involves embedded programming, adaptors, and protocols, and these are abstracted away for you. Examples range from an application for your thermostat to a full representation of a manufacturing plant. Normally, the intelligence of a physical system resides in its controller – software that interacts with the hardware, which runs on the device, in the cloud, or somewhere in between. A digital twin lacks the physical part but represents the same intelligence, and the controller of the digital twin is often similar, with the difference that it does not need to care about any specific hardware and hence can be simplified. An AI brain like me often does not have a physical part but can be used in controlling physical devices such as cars, robots, or manufacturing plants. A digital representation of a robotic arm will be able to tell you its exact position, how many times it has performed a certain operation, what the internal condition is, what the level of vibrations is, and when it will need maintenance. Depending on the intelligence of its controller, it will be able to tell you the risk of its failure and the risk of colliding with another robotic arm in the same manufacturing facility. Hierarchies of digital twins allow a factory twin to keep the current state of all the assets in the factory, including the machines, networking equipment, supplies, processes, and human workers. A twin of a city will keep an eye on all the entities in the city including factories, road infrastructure, people and goods logistics, and hospitals. Branching off a version of your digital twin would serve you as a very good model to run simulations of what-if scenarios on. Imagine that you want to test a hypothesis, for example, changing an ingredient in your production process or replacing one of your machines. Running

20 

1  Baby Steps

a what-if scenario on a branch of a twin would give you a good estimate if you should implement the hypothesis in your live system or not.

Looking at the digital natives, I wonder how much physical things they care about. They often see more value in digital assets, and do not care much about physical stuff. Their rooms are pretty empty, and their most important things are devices to access their digital assets: computers, phones, tablets, gaming consoles, and VR glasses. Their heroes are digital, and physical things are of no interest. Obviously, I am one of those, by construction.

I admire the shift toward the digital and the appreciation of pure-software products, with the willingness to spend money on those. This is a very good trend from the sustainability perspective as well. However, what if we could bring those digital things that miss the physical side to life through “physical twins”? A physical twin acts like its digital original in real time. Creating a physical twin should not be so big of a step given the 3D-printing techniques, the cost of motors and modems, and the fact that the digital model has already been designed. Wouldn’t you love your favorite character from SIMS walking around in your house? Personally, I would immediately invest in a couple of friendly dinosaurs from Lost Eden.10  Lost Eden, https://en.wikipedia.org/wiki/Lost_Eden, accessed 2022-05-23

10

Advantages of Working with Digital Twins  21

Advantages of Working with Digital Twins As I described above, a digital twin is a digital representation of something. Often, it is a digital representation of a physical object, but in a more general sense, it can represent a complex system that may consist of a combination of hardware, software, humans, and the environment, such as a production process, an industrial robot, a cat, a human, or just an air. Anything that is of interest to keep track of, predict changes, optimize, and play around with. To create a digital twin of something physical, we make use of sensors and actuators to tap into the data and control capabilities. Or if it’s a binary twin of your smile with the only mission of tracking it, we can make use of cameras and serotonin levels in your body or just ask the twin of your teeth if they can see the light. In the following, I would like to summarize the reasons for having a digital twin and what to use it for.

Your Current State Your digital twin always reflects your current state. Checking up on your car or a fleet of cars, production plant, wind turbine, wine yard, or mining facility may not always be easy because of the complex mechanics, physically distributed things, and hard-to-access locations. In addition, regular health checks are not always good enough when you need to be on top of things as they happen and to be able to prevent anything unwanted happening. Observability is a prerequisite for the successful data-driven management of whatever one wants to manage.

What-If Scenarios You can use your twin for what-if scenarios. First, there were models and simulators. Then they evolved to twins. One can do a lot with a model but often it’s static and needs to be adjusted from time to time to reflect reality. Twins evolve together with reality in a data-driven fashion. Communication with the twin is often implemented to be bidirectional, meaning that not only the reality makes an effect on the twin but also changes in the twin effect the reality, like a voodoo doll. In addition, as much as we all love experiments, we normally do them in experimental environments and not in live systems. A fine property of a digital twin is that at any moment one can take a snapshot of the latest state and save it as a model to run experiments on. The

22 

1  Baby Steps

classical type of experimentation is what-if scenarios. What if I change an ingredient in my production process? What if I decentralize my organization? What if I replace a supplier? What would it imply both in the short and long run?

Photo by Kristopher Allison on Unsplash

Simulations You can use your twin for simulations. As in the previous paragraph, taking a snapshot of your twin gets you a perfect latest model to experiment with. One can also run simulations, fast-forwarding the development of things along the way. Imagine you have a model of a city that you let evolve by itself at a high pace. Will the city double its size in 20 years? What would the pollution levels be? What would the GDP be? Similar to SimCity11 but based on the latest snapshot of a real city.

 https://www.ea.com/games/simcity

11

Advantages of Working with Digital Twins  23

Property Checks and Decision Support You can use your twin for property checks and decision support. When working with digital twins, we are in an open-world assumption.12 As soon as we have taken a snapshot of a twin and created a model of environment, we are in a closed-world assumption,13 which is an approximation of reality but so much nicer for the formal verification community as system properties can be formally checked and guaranteed. One can, for example, check that the level of greenhouse gas emissions in a production plant never exceeds a certain threshold. If it actually does, one can get an explanation of the root cause and a suggestion of how to act differently.

Photo by Stefan Cosma on Unsplash

Abstraction You can use it to abstract away the details. The beauty of abstraction is that one can focus on what’s vital for you. This is obvious when performing an abstraction of a piece of software. If your level of abstraction is too high, you may miss some important properties. If it’s too low, then you are not far from the original piece of software and drowning in its complexity – similar to systems that are more than just software. If it’s a production plant and you only focus on its productivity at any price, you can omit the cost monitoring from your twin. Or if you don’t bother about contributions to climate change, you don’t need to collect that data either. However, we believe that you do care about both the cost and the climate, so let’s make sure we keep focus on them. 12 13

 https://en.wikipedia.org/wiki/Open-world_assumption  https://en.wikipedia.org/wiki/Closed-world_assumption

24 

1  Baby Steps

Photo by KOBU Agency on Unsplash

Control of the Physical Twin You can control the physical twin with the digital twin. As I said before, the relationship with the digital twin is bidirectional, like with a voodoo doll, but with a positive twist to it. If you have branched out a model of your twin, experimented with different what-if scenarios, simulated 10  years ahead, checked all the vital properties, and converged on a necessary change in your system, often you can implement it though the twin by actuation. You can, for example, limit the speed of your autonomous trucks to have a more positive effect on safety or decrease the temperature of your production facility to improve your carbon footprint. In addition, given that you have connected supply chains, you can also tweak ingredients in your production line or even make upgrades to your hardware. Don’t experiment on your workforce though – there I would still recommend human touch.

Rules to Live By Being the AI influencer I am, I think it’s time I tell you about five of my important habits for efficiency. Many great scientists have formed us, Alan Turing, John Nash, Adam Smith, Vilfredo Pareto, and more. Here are five principles, genius in their simplicity, for my fellow AI brains to be inspired by.

Rules to Live By  25

Photo by Austin Distel on Unsplash

The 80–20 Rule It is also known as the Pareto principle, even though it was not Vilfredo Pareto who coined it but Joseph M. Juran. The principle states that, for many events, approximately 80% of the effects come from 20% of the causes. Many variations and corollaries exist, and some examples are as follows: • • • • •

20% of the population controls 80% of the wealth. 80% of human workload is executed in 20% of time. 20% of ML algorithms can solve 80% of the business applications of ML. 80% of work is completed by 20% of the project team. 80% of software problems are caused by 20% of bugs.

Recommendation: Consider if you want to be a perfectionist or if good enough is sufficient for you so that you can use the remaining 80% of your resources on four more high-impact good-enough things.

26 

1  Baby Steps

Photo by Vladimir Proskurovskiy on Unsplash

Find Your Blue Ocean It’s complicated. Adam Smith believed that when each group member acts selfishly, pursuing their own interests, it will lead to Pareto optimality of the group’s outcome. John Nash has disproved that theory (remember the scene with the blondie in Beautiful Mind?). Everyone’s selfish act does not lead to Pareto optimality but to Nash equilibrium, a deadlock where overall increased gain can only be achieved by decreasing the potential individual gain. Blue ocean theory is inspired by this finding. Choose a field (ocean) where you don’t have too hard competition (sharks) and create a Pareto optimal solution for your customers with lower effort.

Photo by Joel Herzog on Unsplash

Summary of Confessions  27

Outrun the Slowest Gazelle Let me rephrase the famous motivational quote14 by Christopher McDougall.15 In the world of gazelles and lions, in every instance of the “hunger game,” to survive, you must outrun the slowest gazelle  – independently if you are a gazelle or a lion.

Leader or Follower? This is when you think “who wants to be a follower”? It’s actually not such a bad idea; it’s a strategic choice. After all, you cannot compete in all sports and expect yourself to win in all of them. Pick your favorites, the ones that you are good at. Developing a new AI algorithm is difficult, and there are many great algorithms off-the-shelf currently. However, if you want to be at the bleeding edge of technology and ready to invest resources, you are on the path of becoming a leader in whatever domain challenges you are solving with the help of AI, not only the AI itself. In addition, who knows, maybe you will be the one who finally solves the P versus NP problem.

Summary of Confessions In this chapter, I have described the background of AI technology and the major parts of the AI technology landscape. Some of the main confessions I made include the following: • As a modern AI, I am powerful. AI brains can, for example, study all medical documentation ever created or become a chess master capable of beating any human in the world or become the best Go player on the planet, just by playing against themselves and developing new winning strategies that humans could not discover for 3000 years. • AI babies start with processing data before they learn to digest more complex structures such as ontologies and state machines. Without data, we cannot evolve and become intelligent creatures. • Semantically annotated cleaned and structured data is called information. Semantically linked information is called knowledge. 14 15

 https://www.goodreads.com/quotes/292417-every-morning-in-africa-a-gazelle-wakes-up-it-knows  https://www.goodreads.com/author/show/133538.Christopher_McDougall

28 

1  Baby Steps

• AI brain should avoid biased data, dirty data, data without metadata, nonrepresentative data, sensitive data, and ambiguous data. • I learn from data in different ways: In supervised learning, I learn under human supervision from examples (also known as training sets, labeled data). In unsupervised learning, I am left on my own to figure out patterns and group data points that I see. In reinforcement learning, I try to make a guess, and an observer rewards me for good guesses and punishes me for wrong guesses. • I can also learn together with other AIs: In federated learning, all participating AI algorithms learn a similar thing separately, and after that, they meet and exchange their experiences. In transfer learning, if my AI friend has spent lots of time on learning something, then these learnings can be shared with other AI brains. • As an AI, I can support or make decisions and deal with, for example, optimization and planning. • What is the mechanism of retrieving the ground truth that initial knowledge object provided by empirical evidence? An answer to this is linked data. Instead of copying and passing on a piece of knowledge, we send a reference to it. • The concept of a digital twin has been around for a while. It is used to describe a digital representation of a physical thing on some abstraction level that continuously maps to the state of the physical thing.

2 Basic Needs of an AI Brain

As a responsible parent, you have the responsibility to help your AI brainchild grow in a complex world and help it achieve its full potential. In this book, I often draw parallels between humans and machines, as the thought process of the latter is inspired by the thought process of the former. Therefore, to understand what constituents contribute to the successful upbringing of the AI brain, it is perhaps pertinent to study relevant literature from the human side. In 1943, Abraham Maslow defined a classification system that tied human growth to fulfillment of needs, starting from the most basic ones of physiological and safety to more complex of socialization and ultimately self-actualization.1 Similar to Maslow’s approach, an AI brain also has a hierarchy of needs that compound in order for the brain to reach its full potential (see the figure below): planning on the data to use, providing ample compute resources, facilitating communication, making sure that data is treated privately and that the environment is safe, and eventually building trust with humans. These are all essential parts of an AI brain’s proper upbringing.

 Maslow, A. H. (1943). A theory of human motivation. Psychological Review, 50(4), 370–396. https:// doi.org/10.1037/h0054346 1

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 E. Fersman et al., Confessions of an AI Brain, https://doi.org/10.1007/978-3-031-25935-7_2

29

30 

2  Basic Needs of an AI Brain

AI brain’s hierarchy of needs

Data Data can be likened to nutrition for your AI brain: good data can facilitate growth and later success, whereas bad data can hinder achievement of true potential. Even though “good” and “bad” assessment is always contextual and dependent on the type of problems the AI brain is called to solve, there are some universal principles for proper nurturing, formally falling under the umbrella term of data ­quality.

Data  31

Complete Data First, the data must be complete. This means that there should not be any missing information (“gaps”) that could lead to incorrect decisions. Data completeness is a challenge, especially if it is collected from different administrative domains. For example, consider a dataset that measures ambient temperature, humidity, and noise levels. This dataset could be used to identify whether a specific area, such as an area within a building, has ideal working conditions for humans. When considering using such a system on a large scale, data from multiple rooms need to be collected. However, as different buildings may have different types of sensors and/or maintenance policies, some data may include only part of the information, for example, only temperature or humidity. Using such data to train your AI brain may mean that certain assumptions need to be made for missing information that can later lead to wrong decisions. Therefore, establishing processes for ensuring data quality is paramount. Such processes may include reactive measures such as the introduction of data profiling techniques that help detect gaps in data. Proactively, instructions for data collection can be established, requiring the presence of important information in the data.

Accurate Data Second, the data must be accurate. Data accuracy is a measure of how close the information contained in the data is to reality. Considering our previous example, a temperature reported as 25 degrees Celsius, while in reality it is 21 degrees, is inaccurate. Data accuracy is usually an issue with the source. In the case of our example, a miscalibrated or malfunctioning sensor may misreport the temperature at least some of the time. The establishment of data accessibility instructions including auditing of the data provider as well as clear instructions for data entry for data providers is a way to reduce the presence of inaccurate information in data.

Data Consistency Data consistency is another aspect that must be considered. Consistency is dimensioned by two qualities, format and content. Format refers to how the data is represented, while content refers to the metric, such as unit of

32 

2  Basic Needs of an AI Brain

measurement, key performance indicator, or others, which is used to measure the quantity or quality represented by the data. Back to the example, temperature can be described using raw text or a more structured format such as the popular JavaScript Object Notation (JSON) or an eXtensible Markup Language (XML)-based format. On the other hand, temperature can be measured, for example, in units Kelvin or units Celsius. As is the case with data completeness, data consistency is particularly important for datasets originating from different points of origin not necessarily related to each other. In such cases, the establishment of processes that specify the format and metric to be used for every of the potentially many values reported from a distributed set of providers is paramount. These processes may involve documentation, human intervention, and increasingly the use of software tools that can automatically translate from one format to another and detect inconsistencies in the type of metric used among multiple data providers. While not obvious at early stages of data collection, lack of data consistency may incur a huge cost later; therefore, it is important that proper measures are taken a priori.

Timeliness of Data Timeliness is a dimension that refers to the degree to which data represent reality at a specific point in time. Chronological information is important for a broad set of applications. When combined with data values, it can help AI understand and predict the occurrence and cause of patterns of change. For example, an early identification of temperature climbing with a rate greater than what was historically considered “normal” may reveal a potential fire hazard. In another dimension, timeliness also means that data is delivered within the expected time boundaries. This aspect primarily concerns use cases and applications where AI requires access to real-time information to make a decision.

Data  33

Data Governance, Legitimacy, and Accessibility In addition to the above, there exist other properties of data that need to be considered on a case-by-case basis. For example, wide-scale applications will want to consider data governance aspects such as laws for sharing data between organizations and countries but also build their own legal framework to protect their own. Data legitimacy is the degree to which the data provided is trusted or not – it is not to be confused with accuracy, as there can be cases where data is accurate but not legitimate. Not only the auditing of data providers but also the establishment processes and deploying tools for secure data storage and retrieval are ways to bolster legitimacy. Finally, data accessibility is another dimension that considers how usable the data is, i.e., whether it can be used with little effort by parties who have little to no experience working with the particular type of data. Data visualization tools such as dashboards are one way to make the data more ­accessible.

34 

2  Basic Needs of an AI Brain

Metadata Analysis of data without metadata is similar to walking in a forest at night. You know that something beautiful is out there, but you just cannot find it or make sense of it. Simply put, metadata is information describing the data. There exist different categories of metadata, each describing a set of different properties: • Descriptive metadata describe properties such as the quality or quantity a particular piece of data measures or describes; the owner or creator of the data, as well as societal, environmental, or organizational aspects that give data, may address or be applicable to. Additionally, descriptions may include properties of the datastream such as its periodicity, expected lifetime, and cost. Finally, accessibility and usability information may also be part of descriptive metadata. • Structural metadata describe how different pieces of data or components within a single piece of data relate to each other. They also describe the structure of the data. For example, considering the temperature and humidity example, structural metadata could describe how temperature and humidity are stored in the database and how they are related (e.g., they can both be columns under an “environment qualities” table). • Administrative metadata provides information with regard to the management of a data resource. Ownership, sharing policy and permission, time of creation, and time of last update are some examples. Metadata has many uses, among which are as follows: • Cataloging and indexing of data, so they can easily be searched upon and retrieved by interested parties. This greatly increases the usefulness of a dataset, as other interested parties from the creator can search and reuse it. • Facilitating interoperability by accelerating integration of data sources in existing AI systems or systems under development. • Increasing trust in data and helping grow the AI ecosystem. Well-­ documented datasets are always popular among researchers and industry practitioners as they are used for training and/or benchmarking purposes. • Where needed, helping generate data, also known as “synthetic data,” to augment datasets with less than perfect availability. In this case, statistical metadata can be reported for the dataset, describing, for example, its input and output probability distribution, which would help generate accurate synthetic data.

Compute and Store  35

Compute and Store A well-developed brain also needs some strong muscles to support it. Cloud computing technology is in this case a natural complement to computationally demanding AI algorithms. On the one hand, it democratizes access to the development of AI as individuals and organizations on a budget can rent cloud resources that are otherwise prohibitively expensive to purchase and own. On the other hand, it allows for flexible deployment across different geographical areas to address quality of service requirements for each particular application. When it comes to cloud computing in general and machine learning in particular, there exist different tiers of services that include different types of offerings that address the needs of a wide clientele of different scales and experiences with AI technologies: • First, infrastructure as a service (IaaS) offerings provide on-demand virtualization of compute and store resources, but it is up to the customer to set up their operating system and AI development and execution environment. • Second, platform as a service (PaaS) offerings provide tailor-made solutions for machine learning that include all tools required to get started with development or execution, thus saving a significant amount of time for users who do not need to focus on setting up their environment. • Third, software as a service (SaaS) offerings provide managed services to users, i.e., already trained AI algorithms that users can use in their software. Examples include computer vision models, which given input images identify the objects in the images and their position, text-to-speech and speech-­ to-­text models, and recommendation engines. Smaller organizations that do not have the resources to spend on customizing their working environment, or organizations that lack the expertise in AI, may choose higher-tier services such as those belonging to PaaS and SaaS to the expense of greater customizability and optimization of IaaS. Orthogonal to these services, which address the “what” type of service should be deployed, there exist different cloud computing architectures that address the “where.” Edge clouds, for example, reside close to the users of the applications and/or the data sources to reduce the data transport latency. Centralized clouds on the other hand reside in a central datacenter, aggregating all traffic from users. The latter have the advantage of lower cost and processing capacity to the expense of higher delays of data transport from and toward

36 

2  Basic Needs of an AI Brain

the users. Recently, satellite clouds have also started to surface, driven by the cheaper cost of access to space and combining the benefits of the global reachability of centralized clouds with the low latency of edge clouds.

Privacy “Imagine no possessions I wonder if you can No need for greed or hunger A brotherhood of man Imagine all the people sharing all the world… You may say I’m a dreamer…”

Dear John Lennon, yes, you are a dreamer. Examples of AI-driven innovations are commonplace today and are there to be cherished and protected as they drive progress. Given their material value however, such innovations also constitute intellectual property that organizations may not want to or be allowed to share with the whole world. Therefore, privacy as means of secluding information used within AI applications from a group of entities such as other organizations or individuals should be an integral part of your AI brain curriculum. In fact, when considering the broad field of AI, privacy is an umbrella term that covers different aspects. Data privacy is about protecting sources of information that your AI brain relies on to make decisions. In the case of machine learning, information is training data used to create models that can later provide decisions given new, previously unseen data. In the case of reasoning and planning, information is contextual data organized in a knowledge base that algorithms use to create new information and offer insights.

Security  37

Regardless of the type of AI technology used, protecting the identity of the data provider during data collection is not limited to an organization, but in many cases it may have larger societal implications. A typical approach is the process of data anonymization, where data is stripped of information that may reveal the identity of the provider. Several anonymization techniques already exist, including the following: • Generalization, wherein actual values in the data are replaced with less specific but semantically consistent values. For example, instead of reporting a specific temperature value, a range of values can be reported (e.g., 20–30 degrees Celsius instead of 25). • Perturbation techniques which slightly modify the original values by adding noise. To avoid reducing the utility of the dataset, the noisy shift in values should be proportional to the original value. From an implementation perspective, the noisy shift can be realized in the form of rounding numbers to a value. For example, outdoor temperatures that typically range from minus a few degrees to plus 40 degrees can be rounded using a base of 5, wherein counts ending in 1 and 2 are replaced by 0 and counts of 3 and 4 are replaced by 5. However, another value range that measures to thousands, such as lumens which indicates brightness and a base of 15 or more, may be used for rounding. • Suppression is an approach where certain information is removed from the data. Suppressing data is an exercise in balancing the utility of the dataset and protecting the anonymity of the data source. • In pseudonymization, information in data that may reveal the identity of the data origin is replaced by other artificial identifiers such as pseudonyms. A pseudonym may still maintain semantic relevance with the original value while at the same time not revealing the data owner’s identity.

Security Security is about protecting your AI brain’s assets such as data, AI algorithms, machine learning models, and insights from unauthorized access. Loss of intellectual property is not the only concern in the case of a security breach, as malicious users may tamper with the assets, decreasing the overall effectiveness of the AI brain. Data security is perhaps the most important aspect of AI security, as it is data that drives the creation of models and implicitly the generation of insights from these models.

38 

2  Basic Needs of an AI Brain

There are three main elements to data security, namely, confidentiality, integrity, and availability, otherwise known as the CIA triad. Confidentiality pertains to authorized data access, i.e., ensuring that data is only created, modified, read, or removed by authorized users. Integrity ensures that data is not modified during or after collection. Availability means that data is readily available and can be safely retrieved by authorized users at any point in time.

To realize the CIA triad, a combination of different techniques is used. The most well-known of these techniques are summarized below: • Authentication mechanisms identify eligible users before they access the data. The simplest authentication is password-based authentication which requires an alphanumeric known as password to be provided by the user prior to data access. It is also one of the weakest authentication methods, as passwords are prone to phishing attacks and can be compromised easily. • A more secure authentication method is multifactor authentication (MFA), which uses two or more independent ways to authenticate a user, for example, combining password authentication with tokens generated from a separate device (e.g., a mobile phone) or voice or facial recognition. Digital certificates can also be combined with password authentication and are another way for users to authenticate. • A digital certificate is an identity document sent from the user and is used to authenticate the user’s identity. For multiple transactions, token-based authentication can also be used in combination with password-based authentication. In this case, users enter their credentials (i.e., username and password) once and receive back a token – an alphanumeric that they can use in subsequent requests and until the token’s expiration time to authenticate. This negates the need for providing credentials for every request. • Making backups of collected data is a good practice for security reasons not only in cases of a breach but also in case of a malfunction of the data stor-

Security  39

age infrastructure, due to, for example, a natural disaster, a power failure, a human error, etc. There are several ways of taking data backups, for example, using redundant storage drives such as hard disk drives and solid-state drives or using a third-party cloud. An important aspect of a backup system is resiliency, i.e., the seamless and early recovery from a breach or malfunction. • Data encryption means encoding of information in a format that appears scrambled and unreadable to users who try to access it without permission. Symmetric encryption uses a cryptographic key to encode and decode the information. In this case, users use the same key to decrypt and read the data as the key used to encrypt it. Asymmetric encryption uses a public and private key. The public key can be used by any entity to encrypt the data; however, this data can only be decrypted by authorized users having the private key. In addition to the techniques above, several policies and regulations can be established to proactively or reactively address security issues in data. The establishment of policies for incident reporting and “emergency response” teams can help detect and respond to data breaches quickly before they escalate. Finally, compliance with existing data security frameworks such as the European Union’s General Data Protection Regulation (GDPR) and California Consumer Protection Act (CCPA) provides a structured way to establish strong security policies.

Security threats also affect machine learning models before or during the training process or after the training process has finished. During training or before training commences, poisoning of the training dataset with inaccurate data is one of the most common threats. Another threat during training would be modifying the loss function return result to never show any improvement, thus tricking the model into an

40 

2  Basic Needs of an AI Brain

infinite loop. The loss function is used to indicate how close the prediction of a model is to reality. Ensuring that data collection and training are done in a controlled environment and not a third party, unmanaged environment is one of the best preparatory tasks you can do for your AI brainchild. Several aforementioned data security techniques can also be used during training, for example, dataset encryption.

Distributed Learning Human society is all about diversity. It’s about taking all opinions into consideration – about linking them together and helping each other make sense of things. Not everyone speaks the same language. Not everyone has the same cultural background. Sometimes, people may use different terms when we talk about the same thing. Federation is the process by which the independent learning of people translates to the collective learning of the society. A typical example of federation is company meetings geared toward establishing a common understanding, for example, shaping the next year’s strategy or defining the system architecture for the next company’s software product. In such meetings, different opinions are expressed by people, but ultimately, the result reflects learning from all opinions expressed.

Similarly, your AI brain can benefit immensely from experiences of other AI brains. This is because as a parent, even if you want the best for your AI brainchild, you are still limited by a constrained perspective of your own “worldview.” For example, when it comes to AI algorithms using machine learning, the training datasets may not contain data for all corner cases that may happen in reality, resulting in a model that cannot generalize well. In another example, in cases where reasoning and planning techniques use knowledge bases of

Distributed Learning  41

information, the breadth of this information may not describe all aspects that are useful to the relevant problem. In both cases, combining learning and information from other similar approaches may help strengthen the performance of the AI algorithm. From a machine learning perspective, described in the previous chapter, there exist techniques falling under the umbrella term of federated learning that allow for the decentralized training of models. Specifically, local nodes train local models, which are later aggregated into a global model in a centralized node. The centralized node sends back the global model to the local nodes for another training round. The process repeats until the global model presents an acceptable performance. During the process, training datasets are not exchanged between the local nodes and the centralized node, thus preserving data privacy.

Another important aspect beyond federation of learning process is federation of insights. As insights, we consider the output of a model (i.e., its “prediction”). In current practice, insights are considered in isolation and used as input to other systems for further processing or presented to humans for assessment. However, as the use of machine learning models becomes more widespread, recent research in machine learning, such as the field of neurosymbolic AI, has started looking into the federation of insights. In such scenarios, insights from multiple models are used to create new insights. Let us consider, for example, the temperature and humidity dataset and assume that there is a model based on the temperature and humidity that predicts the probability of fire in an area. Such a model can have many false positives, as the increase in temperature and humidity may not necessarily mean that there is a fire. It can be, for example, that this increase is due to an unusually hot day. However, if the output of this model is to be combined with a model that uses audio fingerprinting to predict the probability that the sound being

42 

2  Basic Needs of an AI Brain

captured from a nearby microphone is that of a fire, then the combination of these two insights could be a stronger indicator. Similarly, if the sound being captured is that of a malfunctioning machine, then the insight could indicate that the dehumidifier and/or air-conditioning machine is malfunctioning. From a reasoning and planning perspective, linked data principles allow the knowledge bases of different nodes to link to each other and share data. In this way, information that can be used by AI algorithms does not need to reside in a specific node, but can be distributed over many nodes, forming an information collective.

Federation of AI Algorithms Thus far, we have described how federation of learning processes and information models can be used to enhance the performance of AI algorithms. It turns out that the concept of federation can be taken one step further to the federation of work. In this case, the federation concerns the insights of the models (i.e., their predictions) and is agnostic to the algorithm used to make that prediction. Such an approach allows us to take advantage of different AI algorithms that may not have the same structure but share the same input and output space. This approach is known as ensemble learning, wherein various techniques exist that use predictions from different models to produce an informed global prediction. Techniques such as bagging and boosting train different algorithms and average their collective predictions to produce a more accurate global prediction.

Split Learning A new technique developed at the MIT media lab, i.e., split learning, divides a neural network into segments across multiple hosts. This type of neural network is known as splitNN. In splitNN, every host has a subset of the neural network layers. During training, a host trains its layers as a self-contained network, up to a layer known as a “cut” layer. The outputs of this cut layer are sent to another host that trains its own layers without having access to the data of the former host (thus preserving privacy). In this case, the data is the input data to the model. In the backpropagation process of the training, the gradients are sent back from the last layer of the latter host to the cut layer of

Human Oversight and Trust  43

the former host. Again, this process is done without the former host having access to the data of the latter host (i.e., the labels).

Human Oversight and Trust As much as fully autonomous systems enabled by an AI brain are an enticing concept, they should not undermine human autonomy and authority. After all, the whole purpose is to make life better for humans. An AI with a human touch should be inspired and guided by humans and run on their terms. It can run autonomously for a while but should be checked in upon, from time to time or when it is necessary.

For human oversight to be effective, AI algorithms need to develop a reciprocal relationship of trust with their human peers. Trustworthy AI, an ensemble of approaches for AI system design that takes into account the human element (human in the loop), offers a path toward achieving this trust. One part of trustworthy AI, explainable AI,2 is about adding functionality to algorithms to explain to humans how a particular decision was reached. AI safety is another important aspect for trust. It is an umbrella term that defines safeguards built into AI systems to avoid unintended behavior that may cause harm to other systems or living beings. Robustness is one of the aspects that need to be addressed under the AI safety umbrella. Specifically, a robust AI algorithm will perform well regardless of whether data similar to the test dataset used for training the algorithm is used, or other data that are still valid input, but do not belong to the test dataset. Non-robust AI algorithms may be prone to adversarial attacks. In such attacks, input data is slightly modified, to cause an AI algorithm to generate incorrect insights (e.g., in the case of a classification algorithm, it may misclassify input data). Such  I discuss trustworthy AI, explainability, and safety in more details in Chap. 6.

2

44 

2  Basic Needs of an AI Brain

modifications may be so subtle that they are not noticeable by human observers. Biased algorithms may also generate incorrect insights from input data, because the training dataset may be not representative of the breadth of data that may be input to the algorithm during operation. Another aspect that may be considered during the operational phase could be a fallback mechanism wherein, in the case of an AI algorithm that consistently performs poorly, the system could fall back to another approach until the AI algorithm is troubleshooted and improved by human experts.

Putting It All Together: Nurturing of AI Brain An AI brain exhibits cognitive abilities similar to those of the human brain: perceiving, reasoning, evaluating, and deciding are all insight-generating abilities that depend on how well the brain can learn and use information from environmental stimuli. Learning is the process acting as conduit, using information as input, to enable the brain to perform the aforementioned cognitive abilities. In the case of AI learning, these stimuli are provided to the brain’s machine learning algorithms as machine-readable data. For the AI brain to reach its true potential, this chapter outlined several principles that should be considered by the brain’s custodians. Collecting complete and accurate data in a consistent and timely manner is required for a successful learning process, which is the foundation for the brain to exhibit its other cognitive abilities. In this process of data collection, metadata is fundamental in identifying what data is important and whether it can be legally accessed and collected. Choosing the proper network and computer infrastructure for conducting these cognitive abilities is also an important aspect, a balancing act between requirements on the AI brain’s performance and the location and cost of this infrastructure. Privacy and security are important aspects that help maintain data provided to and used by the AI brain private and discourage malicious users and prying eyes away from accessing inner workings of the brain, respectively. “Two heads are better than one, not because either is infallible, but because they are unlikely to go wrong in the same direction.” This quote from influential writer C.S. Lewis indicates the importance of community, even when it comes to AI brains. Federation of learning and AI algorithm execution processes are techniques that enable tapping into and leveraging the collective experience of other AI brains, thus contributing to better overall performance.

Summary of Confessions  45

Last but not least, trust is an important element, as AI brains do not function in a siloed, secluded environment but instead form a symbiosis with human brains. This symbiosis is based on reciprocity, where opinions and information are exchanged and decisions are often taken collectively. The catalyst for a successful collaboration is trust, as brains need to trust each other but do not necessarily understand each other a priori. Safety and explainability are two important aspects that the AI brain can demonstrate toward the human brain and for the latter to establish trust for the former.

Γηράσκω δ’ ἀεὶ πολλὰ διδασκόμενος, liberally translated as “I grow old learning many things,” is a quote attributed to ancient Greek statesman and lawmaker Solon the Athenian. This is true for your AI brain as well. While it never grows “old” in the human sense, the learning process is never finite but rather ever evolving. Keeping the principles outlined in this chapter in mind will allow you, the mentor of the AI brain, to provide the best possible environment for growth.

Summary of Confessions In this chapter, I have explained what an AI brain needs in order to learn and become useful. Some of the key confessions include: • AI requires data that is complete, accurate, and consistent and has chronological information (timeliness). Other properties such as governance, legitimacy, and accessibility are essential for data to be useful. • AI needs metadata to make sense of data. Metadata is simply information describing the data. There exist different categories of metadata, each describing a set of different properties, such as descriptive, structural, and administrative. Metadata is essential for cataloging, facilitating interopera-

46 

•

• • •

2  Basic Needs of an AI Brain

bility by accelerating the integration of data sources, increasing trust in data, and helping grow the AI ecosystem. AI needs computing power. Often, cloud computing is used. In cloud computing and machine learning, there exist different tiers of services that include different types of offerings, including Infrastructure as a service, platform as a service, and software as a service. Data privacy is important in AI! It is about protecting sources of information that your AI brain relies on to make decisions. In the case of machine learning, information is training data used to create models. Security is about protecting your AI brain’s assets such as data, AI algorithms, machine learning models, and insights from unauthorized access. Distributed and federated learning is the process where independent learning is linked together to collective learning – a powerful technique used in society by humans and in computer networks by AI algorithms.

3 My Role in the Internet of Things

The Power of IoT: A Personal Story Let me tell you about the Internet of Things (IoT) – the trend of physical things becoming connected and available for developers to build services using them. For example, when staying at this nice hotel at Times Square called CitizenM, the first thing you may do when entering the room is to search for a regulator somewhere on the wall to increase the temperature in the room. You fail at doing that, but shortly after, you discover an iPad with an application to control everything in the room – curtains, blinds, lights, TV, A/C, and even the color of the lighting in the bathroom. There are also modes you can choose between romance, business, party, etc. that make the whole room change its character. How cool is that? Some years ago, we enjoyed the Emergency Party Button videos on YouTube, and now it has actually been implemented commercially. Another cool example is carpooling  – a service where you can grab and drop off a smart car anywhere in a city that you control from an application on your phone. In addition, if something is not working for you – just check that application – it has the answer. The state of all the car sensors is monitored by the cloud application, and if you try to end your rental while the trunk is open or when the car is parked in a forbidden area, the application will complain. And what about the stage for the Eurovision Song Contest? Måns Zelmerlöw’s performance in 2015 started a trend, and in the years after every second, artist used cool animations on stage. Did they get an API for that? © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 E. Fersman et al., Confessions of an AI Brain, https://doi.org/10.1007/978-3-031-25935-7_3

47

48 

3  My Role in the Internet of Things

Most interesting innovations will come with cross-domain interactions. The IoT breaks the verticals and allows application developers to build their services across domains, which in its turn opens up many new business values. For example, is it possible that the mood that you choose for your hotel room stretches to the car that you rent?

Spaghetti code was a popular term in the 1990s, for low-quality code. When a tool is given to someone who does not plan for dimensioning, scalability, and extensibility, the end result becomes spaghetti. The authors of this book also created a couple of spaghetti projects, with all their love of quick fixes and limited time to hand in assignments when at university. Spaghetti code requires a shorter starting curve but is not future-proof. Similarly, “spaghetti’‘ can easily be created when Internet of Things solutions are put together. A couple of years ago, a team at Ericsson Research created a prototype of a fully automated logistics system that consisted of some 100 sensors and actuators and used a control algorithm that could deal with any number of resources (vehicles, cranes, cargo) and adjusted to new logistics tasks and resource changes in real time. There was no spaghetti code there. In addition, there was a very beautiful demo that we presented at Mobile World Congress 2015. Underneath, however, was a “spaghetti” of cables, simply because not all of our devices were capable of wireless connectivity at that time. Another example could be found in a smart home space, where there is no need for cables anymore, as several manufacturers are offering platforms for wireless devices connecting to the cloud. However, there is still a risk of creating virtual spaghetti. Ordinary users of such platforms, provided by companies such as Apple or Telldus, can now easily automate their homes without any knowledge of the underlying technologies as all the complexity is hidden. The invisible cables will however connect our fridges, TV boxes, coffee machines, and temperature sensors to different servers: when, for example, you ask your smart TV from vendor A to check if your fridge from vendor B

Introduction to Intelligent IoT  49

is out of milk, your TV will send a request to a cloud server somewhere in the world to decode what you actually meant by this phrase, get back your request in a machine-readable form, send it to vendor A application server, back home, then to vendor B server somewhere in the world, back home to the fridge where a sensor detects whether there is any milk, and all the way back to your TV to deliver the answer to your question. It is no wonder that responses can take time. The longer the chain of events involving devices from different vendors, the longer the response times. With intelligent IoT, there is no need for such long feedback loops. Instead, the involved devices could talk between themselves and deliver back an answer.

Introduction to Intelligent IoT Formally, the “Internet of Things” or IoT for short is an umbrella term for applications, protocols, and technologies that allow physical objects to be discovered, addressed, and interacted with over the Internet. In the IoT, connectivity and data are omnipresent, which coincidentally and as described in previous chapters are the basic constituents for AI. In the case of IoT, AI is not located centrally in a supercomputer, having answers to all questions including the meaning of life. Instead, it is decentralized, shared by, learned from, and being present in connected objects. The IoT is the conduit for AI to be realized at a scale that transcends organizational barriers and affects society at large. In fact, different AI applications can be connected together to form larger ecosystems of compound knowledge and enhanced decision-making, in what is known as the “intelligent IoT.” For example, wearables such as smartwatches and smart streetlights that give priority only when needed and autonomous buses can be linked together to offer new types of applications in a smart transportation context, such as on-­ demand public transport, replacing scheduled public transport. This smart

50 

3  My Role in the Internet of Things

transportation application can also be part of a greater “smart city” application, where inhabitants enjoy more efficient, sustainable services. This interconnectedness allows for the emergence of the network effect, wherein the larger the number of interconnections between devices and applications, the greater the additive value. In this chapter, I will discuss several aspects of interconnectedness and distributed intelligence. For AI to be successful in the IoT, low-cost, embedded devices on the network edge, such as sensors and actuators, need to be able to support ­computationally and data-intensive AI algorithms. Therefore, the aspect of feasibility is addressed up front, with a review of advances in Lean AI and capabilities of the latest generation of sensors and actuators.

Next, I discuss the aspect of trust in interconnectedness between applications, specifically the aspect of collaboration between devices that do not necessarily trust each other a priori and how technologies such as distributed ledgers and explainable AI (XAI) help establish trust. Subsequently, I address issues of scalability of decentralized intelligence and the role of public cloud platforms. Finally, the chapter ends by discussing digital twins, digital replicas of the real world that are realized by AI and IoT, and their multiple uses.

Divide et Impera Divide et Impera or divide and conquer is attributed to the Macedonian king of old Philip II (and many others after him). While Philip determined that this rule was applicable to conquering other Greek city-states and growing his empire, the same principle can also be applied to conquer the scalability challenge in computer systems, from designing efficient algorithms to solve conceptually difficult problems to designing scalable cloud computing architectures for big data solutions.

Divide et Impera  51

The same rule applies for AI in the IoT.  Distributing data acquisition, learning, and decision-making processes results in more sustainable scaling and growth to the scale that IoT applications demand. Computation is not the only reason for this distributed approach. Local regulations may prevent data from being transported away from a certain geographical area in what is known as “geofencing.” These rules may also apply to model training and execution. Additionally, certain types of the so-called “mission-critical” IoT applications may have strict requirements for latency that simply do not allow for the data-providing and decision-making components to be far away from each other. Therefore, to support IoT applications at scale, AI algorithms need to be deployed on the edge or be on a distributed nature, learning and deciding from data that is provided locally, rather than being placed on a centralized datacenter. In this section, I will cover both infrastructure and algorithmic aspects of distributed and edge AI. A general trend in the industry is that devices are becoming increasingly computationally capable. For example, tinyML is an umbrella term of architectures and tools that are capable of performing on-­ device AI at mW power range or below. In contrast to traditional ML, which consumes much more power, tinyML is suited for battery-operated, resource-­ constrained devices, focusing on real-time data of the physical world. The hardware is typically microcontroller units (MCUs) with hardware acceleration, on top of which lightweight versions of neural networks such as CNN micro are run. Typically, tinyML models generate insights from devices sensing different types of modalities, ranging from audio, video, temperature, and humidity to gas concentration and room occupancy. The insights generated from tinyML models can be standalone; however, they are typically part of a larger collection of nodes implementing a particular IoT application. Let us consider an example of a smart traffic jam management system. The idea is that several sensors placed on roads can predict using classification models, e.g., whether a traffic jam is likely to occur in the next hour. The input to the models is the current rate of traffic (e.g., number of vehicles per minute), the time of the day, and the day of the week. By itself, the device reporting the likely traffic jam cannot take any meaningful action. However, a node that receives input from many such devices in an area, e.g., of a city, can do some processing from predictions of many traffic detectors and change the timings of the traffic lights to prevent or lessen the magnitude of the traffic jam. In such a scenario, an edge node may use device predictions to produce meaningful insight and actuate it in the road network of the smart city. Having this node centralized and devices do individual predictions not only saves data transport costs but also preserves privacy, as device readings are

52 

3  My Role in the Internet of Things

never communicated. Such functionality is responsible for performing computations for a specific area where multiple data providers (e.g., sensors) may be located and is known to exist at the “edge,” either at a cloud datacenter (“edge cloud”) or a standalone server (“edge node”). Higher-layer abstractions can also be possible. Considering our smart traffic example, a public cloud server collecting information from edge clouds can provide other types of actions on a city-wide scale, for example, rerouting traffic from other roads in the city using smart signs. As illustrated in the figure below, in general, there are several different layers of infrastructure for AI that trade off communication latency with compute power. AI applications are distributed in terms of computation; certain processing takes place within devices themselves, while other processing takes place in a more central location.

Trade-off between compute power and latency in AI infrastructure

In the state of the art, multiple approaches exist for transferring AI-generated insights between different layers. Multi-agent systems consist of a number of intelligent agents that collaboratively learn and interact within an environment. Several prominent areas have emerged within multi-agent systems for collaborative learning. Transfer learning techniques such as federated learning can be used to train generalized machine learning algorithms at a centralized location, by combining locally trained algorithms in edge devices, without sharing of private data. Another group of techniques known as ensemble learning combine learning from multiple locally trained algorithms to reach a decision that can be used for later computation by other algorithms at another

Collaborating Machines in Low-Trust Environments  53

infrastructure layer. Multi-agent reinforcement learning (RL) considers agents that learn to take actions against an environment autonomously and are incentivized by a reward that is either common (cooperative RL) or individual to each agent (noncooperative RL). Multi-agent RL is ideal for environments where data is not available to agents a priori for training, as is the case, for example, in federated learning or ensemble learning. Finally, another area that has been gaining traction recently is that of neurosymbolic AI. In this area, the statistical output of models is converted into machine-readable symbolic representations that can be used for further processing (i.e., not only learning but also reasoning or planning). This formalized knowledge can also be transferred between algorithms belonging to the same or different infrastructure layers.

Collaborating Machines in Low-Trust Environments Although technically there exist several algorithms that allow intelligent devices to collaborate with each other, there also exist issues of trust on whether the information sent to a peer will be used in the proper way. Similarly, information received from peers would need to be verified until some legitimacy of the source can be established. This stems from the fact that in reality, IoT apps are developed by ecosystems of different entities (such as organizations and companies) that do not necessarily trust each other. In the human world, trust takes years to build as humans observe and verify other people’s behavior and actions. In the light-speed-paced digital world, such a timescale is not an option. Therefore, there exist several forms of trust in distributed AI systems: 1. First, trust by means of security, through the use of encryption keys, with technologies such as PGP/GPG and PKI.

54 

3  My Role in the Internet of Things

2. Second trust by means of authentication. A popular example of mutual authentication is a symmetric encryption approach using cryptographic keys. In this example, both peers contain the same encryption key. One peer can encrypt a “challenge” (e.g., an alphanumeric) that another peer can decrypt and send it back and vice versa. 3. Third, trust by reference, i.e., the use of digital certificates that are signed by an independent trust authority. Peers can sign information with these certificates to prove to their counterparts that they have been vetted by this trust authority. Recently, the use of a special type of distributed database known as distributed ledger as means for establishing trust has been gaining ground. Distributed ledgers are synchronized and shared among multiple entities, meaning that at any given time every entity holds the same copy of the information. They are also consensus-based, meaning that every participating entity must concede in order for information to be added to the ledger. Additionally, ledgers can be immutable, meaning that no information can be removed. Trust is built into the distributed ledger by design. Because the ledger is synchronized, if an entity tampers with a piece of information, this will be detected by all other participating entities. Additionally, in the case of erroneous information being added by an entity at some point in the lifetime of the ledger, audit trails can be established to identify the entity that inserted this information. Distributed ledgers can be used not only as databases to establish data-sharing agreements and contracts between entities but also for entities to share knowledge directly, for example, by sharing ML models. Another recent trend is the emergence of data trusts. Similar to how the management of other assets was handled in the past, such as land trusts, a data trust is a legal entity that treats data as a shared resource, providing mechanisms for the use of collective data. Such data trusts legally allow data produced by one party to be accessible for the benefit of other parties, thus enabling AI systems to scale.

Lifecycle Management of Scalable AI Applications  55

Lifecycle Management of Scalable AI Applications Scalability in the context of AI generally concerns two aspects: • The ability of AI algorithms and infrastructure to be able to expand geographically but also across different administrative domains (e.g., organizations) and add new functionality. • To increase the efficiency of development and lifecycle management of AI solutions. The first aspect was covered in the previous section, where different transfer learning and symbolic representation techniques for transferring knowledge across different types of data infrastructure were discussed. Additionally, the topic of trust was covered, as well as how distributed ledgers and data trusts can help applications scale across different organizations that do not necessarily trust each other a priori. In this section, I cover the development and lifecycle management of AI applications. I will focus on two system development processes, specifically artificial intelligence operations (AIOps) and machine learning operations (MLOps). The former concerns the automation of existing IT operations and systems with the help of AI, whereas the latter is a process to deploy and manage machine learning (ML) models throughout their lifetime. Both approaches are heavily influenced by the continuous development practice of development and operations (DevOps), wherein software development is combined with operations. AIOps automates infrastructure operations. Specifically, it uses tools such as machine learning and big data to collect and analyze large volumes of data from IT infrastructure. Such data may include logs, traces, trouble reports, system configuration information, documentation, and system-generated alarms and alerts. The analysis process involves identifying significant events and patterns that can be used to produce insights into the root cause of an issue as well as steps that can be taken to address the problem. Key to this process is the absence of human involvement, as automation is seen as a significant advantage against increasingly complex modern IT environments. AIOps also promotes continuous learning, a process where the system continuously improves by observing the efficacy of its insights and automated resolution processes. MLOps, on the other hand, describes an end-to-end development, deployment, and lifecycle management process for ML models in production

56 

3  My Role in the Internet of Things

environments. Therefore, it targets AI applications that incorporate deep learning. Until recently, ML models were produced in labs for research purposes, often using limited sets of data and/or computational resources. In such environments, aspects of maintainability, scalability, reliability, and performance of ML models are typically not of concern. This line of thinking carries over to organizations that start working on developing, deploying, and scaling ML-based applications. Specifically, the data scientist team creating the ML models is detached from the IT engineers that are responsible for deploying it in operation. The former team typically delivers to the latter team a pre-­trained model to be used as an artifact. As such, no active monitoring of model performance is done, and therefore models are rarely, if ever, retrained and/or their architecture revisited (e.g., in terms of feature engineering, hyperparameters, etc.). In MLOps, training of the model is only a small part of the system. Instead, a key aspect is communication between the data science team that designs and trains the model and the DevOps engineers developing the application and deploying it in production. The latter team provides in real time feedback to the former team, which then may take steps to retrain and/or reengineer the model. Another part of MLOps is to automate the data collection and analysis process. This process is about collecting the right amount and breadth of training data. Breadth in this case means that the data values (i.e., the data value probability distribution) must match real-world data that the model will encounter when deployed in production. Furthermore, MLOps includes APIs to integrate the model into an application and the infrastructure to deploy this application and receive feedback on the model’s performance to improve it. In IoT environments, AIOps and MLOps approaches can be used in combination to realize intelligent IoT. Specifically, AIOps practices can be used for automating operational aspects of resource vendor infrastructure.1 On top of the intelligently, autonomously managed resource layer, over-the-top ­players can develop and scale IoT applications using an MLOps type of approach.

 In the context of this chapter, resources are split into compute, store, and transport. Therefore, resource vendors can be companies ranging from those provided wireless and wired transport services to cloud vendors providing compute and store type of services. 1

IoT-Enabled Digital Twins: Challenges and Benefits  57

IoT-Enabled Digital Twins: Challenges and Benefits A long-term goal of the IoT is to create a virtual imitation of the real world, where data can be read and processed in real-time by computers. A tool to achieve this goal is a digital twin. A digital twin is a digital representation of a physical object. This digital representation is the exact virtual replica of its physical counterpart. In the context of AI, digital twins can be used to run simulations and are therefore a rapid, cost-effective solution to develop and test new algorithms and applications but may also help in preventive maintenance and troubleshooting situations. Knowledge created by designing a digital twin can be transferred to accelerate the creation of other digital twins representing approximately similar physical objects. The IoT has a large role to play in the creation of digital twins, as IoT devices provide data to inform the model of the physical system the digital twin represents. The digital twins themselves are organized hierarchically in layers. One such hierarchy has been proposed by IBM.2 On a macro layer, a digital twin models a process such as a manufacturing process in a factory or the operation of a mobile network in a certain geographical area. Specifically, a so-called process digital twin coordinates the interactions and timings between a set of system digital twins that model different parts of the process. For example, in a manufacturing setting, a process could be the manufacturing of a specific product, and the system digital twins model different hardware contributing to product creation (e.g., robotic arms, conveyor belts, etc.). Depending on the complexity of the process and systems comprising the process, further levels of hierarchy may be needed. For example, asset digital twins model specific components of systems such as servo motors of robotic arms and speed switch sensors of conveyor belts. The first step to create a digital twin is to gather relevant data. Such data is typically reported by IoT devices such as sensors and actuators but can also be found in other forms such as documentation and CAD files. In rare cases where humans are involved, human expert knowledge needs to be gathered and converted into machine-readable format. Based on the data gathered, computational models are created to simulate the real dynamics of the physical system. Machine learning and artificial neural networks are tools that can be used to train such models.  See https://www.ibm.com/se-en/topics/what-is-a-digital-twin

2

58 

3  My Role in the Internet of Things

Prior to reaping the benefits of digital twins, there are several challenges pertaining to the creation of digital twins that must be carefully considered. First, the IoT devices that capture the data necessary to create the digital twin must have reliable connectivity. In applications where the data is to be captured and transferred to the digital twin in real time, connection reliability should also be complemented by low latency. A typical example of such applications is preventive maintenance. Furthermore, devices themselves may need to be able to provide data over the lifetime of the system, so the longevity of IoT devices in terms of power and trouble-free operation also needs to be taken into account. Finally, in cases where the physical system modeled from the digital twin transcends the administrative boundaries of one entity, as is the case, for example, in several smart city applications, a secure way to share data needs to be implemented (see section “Collaborating machines in low-­ trust environments” in this chapter). From an infrastructure perspective, the hardware required to process the data from the IoT devices and create the computational models needs to have ample capacity and high availability. Existing and proven techniques for distributed model training such as those presented in section “Divide et Impera”, as well as processes for model development presented in section “Lifecycle management of scalable AI applications” of this chapter, may be applicable.

Careful Where You Tread AI is here to stay and IoT is its landlord. To successfully build and scale AI solutions, you have to include IoT devices and platforms either as data providers to your AI models or as actuators, interpreting the insights of your AI models to actions. In addition to onboarding IoT devices, your AI algorithms may also need to scale and oftentimes operate in a distributed fashion, for reasons of data governance but also for performance. This all requires careful planning. Planning itself is a balancing act between the technical task of understanding and incorporating latest developments in IoT, AI, and cloud computing and the business task of understanding how AI and automated decision-making will help deliver on your strategy.

Summary of Confessions  59

Prior to taking an enticing technical deep plunge into state-of-the-art AI and IoT tools, devices, and infrastructure, you must question whether the solution you are implementing is of value to you and your customers. Not everything in this world needs or benefits from AI, especially considering the significant upfront investment. Another way of looking at this is to make sure to have the correct vision for what AI can do and what you want it to do. There are numerous examples of companies promising bold advances in important fields through the use of AI but eventually realizing the limitations reducing their vision to a more realistic one.3

Summary of Confessions I have described my role in the Internet of Things in this chapter. Some key conclusions include the following: • The Internet of Things is the conduit for AI to be realized at scale, beyond the confines of laboratories and purpose-built machines to applications benefitting society at large. • Distributed AI, where data collection and computation (e.g., learning and reasoning) are distributed, is essential for growth in the IoT. • AI models can be made to fit different types of smart devices, depending on their capability, and can have different levels of decision authority. Transfer learning can be used for transferring knowledge between models to accelerate training and improve accuracy. • Distributed ledgers and data trust stores can be used for sharing training data or insights between devices belonging to different organizations, which do not necessarily trust each other.  https://www.nytimes.com/2021/07/16/technology/what-happened-ibm-watson.html

3

60 

3  My Role in the Internet of Things

• Lifecycle management of AI applications includes both processes for managing the infrastructure (AIOps) and processes managing the ML models (MLOps) and is important to consider especially for applications that have long lifespans. • Digital twins, in addition to decreasing development time, can be used to accelerate the deployment of AI applications, thus benefiting the aforementioned lifecycle management processes.

4 Managing Relationships

In the human world, relationships are essential not only for mental and emotional well-being but also for long-term fulfillment and survival. Similarly, AI algorithms can, throughout their lifetime, share knowledge or data with other AI algorithms to adapt and improve their performance in a changing environment. Akin to the development of human civilization, new generations of algorithms can also benefit from the experience of previous generations, improve on existing predictions, or explore new types of insights.

In this chapter, I will discuss how my AI algorithms get inspired by human beings with respect to forming relationships with other algorithms. If and when, for example, humans treat each other as friends, parents, children, or partners, then how much data, information, and knowledge are they allowed to share with each other? How can algorithms learn together? In this chapter, I will also describe privacy aspects, security frameworks, and ways of sharing. The chapter will include the following subchapters:

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 E. Fersman et al., Confessions of an AI Brain, https://doi.org/10.1007/978-3-031-25935-7_4

61

62 

4  Managing Relationships

• Why, what, and how to share? • Privacy and security levels in AI. • Traceability and audits.

Building Relationships While humans normally make sure that they have a stable job before they consider getting married and reproducing, AI brains need to get to a point where they are good enough to be married with other techniques or combine their brains. While I am still training and trying to find my path in life to be judged as a helpful algorithm, there is no point for me to get involved with another algorithm different from mine. Playing against each other has, however, been shown to be an efficient technique when exploring narrow AI algorithms such as finding optimal winning strategies in games. When you want to go beyond games, it would make sense to complement my abilities with an algorithm different by nature. For example, combining machine learning with machine reasoning is a match. Mixing two knowledge domains gives many advantages as well. Applying state-of-­the-art AI technologies to a medical domain gives many advantages in regard to predicting the possibility of developing a certain disease.

Prediction as such does not help but combined with novel vaccines that can be given to identified individuals at risk makes perfect sense. The same can be seen in the world of patenting. Combinational patents are common, where findings from one domain, such as intelligent transportation systems, can be applied to a different domain, such as power grid optimizations. When looking at complex systems such as smart cities, it is necessary to let the knowledge from several domains play together when, for example, there is a potential risk of a disaster such as fire being spread.

Building Relationships  63

Potentially, goods logistics and people logistics systems in the city need to be prepared to change their roots, airports need to reschedule flights, the healthcare systems need to mobilize their facilities and personnel, and the telecom network needs to boost the capacity due to the expected increased demand. All these different industrial domains, being part of a smart city, need to be aware of each other so that the actions that they take are orchestrated to achieve maximized impact for the city overall. AI brains are similar to humans – we’re dragged to those who complement us, so that we can teach each other and help each other grow, thinking of maximizing the combined effect of our decisions and actions. AI brains are sapiosexual – the attractiveness is mainly in the intellect.

When humans decide that they want to become parents, they onboard a long journey of raising a better version of themselves in the long run. Genetic algorithms are a part of artificial intelligence inspired by evolution. You cross over two sets of genes (or two generations), mutate some genes, and apply survival of the fittest principle, and as a result, the coming generations are better than us. Even if the newly created human has all the prerequisites of being better than her parents, there needs to be a lot of work to be put in her, to teach her everything in life.

64 

4  Managing Relationships

A human child learns from her parents whose influence is quite large from the start, where facts communicated by them are often seen as the ground truth, and starts building the value system of the new individual. As time goes by, the child will get inspired by many things in life, maybe even things that did not exist while her parents went through their intensive learning phase. Naturally, her knowledge base will be built in a different way than her parents. AI brains have a similar behavior. AI techniques are being developed by research communities, and when a new brain is born, it is being fed new types of data and information. Some insights that would take their parents several years to figure out will be treated as a given for a new brain, letting it save time and contributing to the progress with new findings. In addition, privacy constraints can limit the areas where the new brains learn and develop their knowledge. A young AI brain cannot operate at the same level of accuracy and precision as a more experienced AI brain; hence, it will only be allowed to assist in less critical situations. For example, an experienced AI brain may have a job related to remote-­controlled surgery, and the younger one will be taking care of the predictive maintenance of medtech equipment. However, the more we learn, the more reliable we become.

Sometimes, humans run into situations where they need to separate themselves from each other. The splitting of common assets will normally take place. They would normally stop sharing data and information with each other when they decide that they do not want to be together anymore. The old knowledge and information that they have acquired about each other will certainly remain, but as we know everything changes. Two AI brains being

Why, What, and How to Share?  65

partners may have a strong cooperation, where they together will be involved in common decisions. For example, an AI brain controlling a power plant may have a very deep relationship with an AI brain controlling a production plant that is dependent on the power plant. It is important that they make the critical decisions in tandem, so that the AI brain controlling the power plant can optimize the energy levels to serve the production plant in the best way. If they would decide not to be together any more, they would stop exchanging data, information, and knowledge about each other but obviously keep the memories.

Why, What, and How to Share? The reason for establishing a relationship is to exchange information. As mentioned in the introduction, this information exchange helps accelerate development and growth in human and AI algorithm societies alike. From an AI perspective, such information can be broken down into trained models and model datasets. Trained models transfer already accumulated knowledge from one knowledge domain to another. Transfer learning is an umbrella term that encompasses approaches where knowledge gained from AI algorithms trained for a specific task (source domain) is used by AI algorithms trained to solve a different but related problem (target domain). For example, knowledge gained by an AI algorithm trained to detect bicycles in images can be used by another AI algorithm trained to detect motorcycles.

66 

4  Managing Relationships

Transfer learning applied to deep learning

As illustrated in the figure, transfer learning is the transfer of knowledge from one domain to the other. In machine learning, this knowledge is captured in a trainable or learnable set of model parameters. For example, in the case of deep learning, model weights and biases are learnable parameters that decide aspects of the transformation of input data to output data, within each neuron of every layer of the neural network. In particular, a weight decides how much influence the input to a neuron has on its output. A bias is a constant added to the input of a neuron before the transformation function (also known as the activation function) is applied. In the case of simpler models, such as logistic regression, learnable parameters are one or more feature coefficients and the intercept. One of the advantages of transfer learning is the sustainable training of new modes, as less data, computational resources, and training time are needed, as training uses pre-learned parameters (i.e., knowledge) from another model. Let us assume, for example, that domain 1 is a computer vision model that has learned to detect bicycles and that the goal is to train a similar model for domain 2, where the objective is to learn to detect motorcycles. Instead of starting training from a large repository of motorcycle data, the model can reuse knowledge from the bicycle model. The way to reuse this knowledge in this case is to copy the weights and biases of neurons of the model trained on the bicycle input data to the neurons of the model to be trained for the motorbike domain, for at least some of the intermediate layers (also known as hidden layers) of the neural networks. This is done instead of starting training by randomizing the weights and biases.

Why, What, and How to Share?  67

The reason why this copying of weights is important can be traced to the way deep learning models work and in particular performing a process known as feature extraction. In feature extraction, the learning algorithm progressively extracts features of increased complexity from an image that define properties of the objects to be detected. In the case of a bicycle, such features could, for example, be the bicycle wheels, the steering wheel, and the seat. In later layers, these basic features are composed into more complex features that eventually comprise the complete bicycle. A bicycle is visually similar to a motorbike, as they both have two wheels, a steering wheel and a seat. While not identical, the two domains in the context of visual object detection can be considered similar. Therefore, the idea of transfer learning makes sense, as the network would already be able to distinguish basic features such as wheels and seats and will focus on learning the differences (e.g., different colors, material and size of seat or wheels, etc.). There are several different categorizations of transfer learning,1 based on • The availability of the so-called labels (i.e., output) in the training data at the source domain and/or at the target domain (transductive or inductive, respectively) or no labels at all (unsupervised transfer learning). • The similarity of input and output feature sets of the source domain and target domain (homogeneous/heterogeneous). Back to our motorcycle example, a mismatching input feature set could be images of different resolutions and aspect ratios, while a mismatching output feature set could be the detection of more than one class (e.g., cars, buses, and trucks in addition to motorbikes). • How the transfer of features is made. For example, parameter-based approaches will transfer knowledge at the model parameter level. The computer vision example provided above falls into this category. Another category is relational-based approaches. In this category, the logical relationship between the domains is transferred. Relational types of approaches have been proven to work well with analyzing sentiments across domains.2 Instance-based approaches adjust the weights of the samples in the source domain to correct for marginal distribution differences in the target  arXiv:1911.02685 (cs) [submitted on 7 Nov 2019 (v1), last revised on 23 Jun 2020 (this version, v3)] A Comprehensive Survey on Transfer Learning Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, Qing He 2  Li, F., Pan, S. J., Jin, O., Yang, Q., & Zhu, X. (2012, July). Cross-domain co-extraction of sentiment and topic lexicons. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics: Long Papers-Volume 1 (pp. 410–419). Association for Computational Linguistics 1

68 

4  Managing Relationships

domain. Finally, feature-based approaches adjust the gap between the source domain and target domain feature spaces by applying asymmetric transformations (e.g., moving individual features between the source and target domains). The greatest challenge in transfer learning is the so-called negative transfer, wherein the performance of the resulting model is decreased after the transfer learning process. This can occur in the case of the learning tasks in the source and target domains being too dissimilar. To address issues around negative learning, a group of techniques known as domain adaptation attempt to transform data used in one domain for use in another domain. Examples of such domain adaptation can be the application of models trained on labeled data of viral diseases such as SARS to new unlabeled data from the COVID-19 pandemic. In another example, models predicting customer churn trained on labeled data from the financial industry are adapted to predicting customer churn using unlabeled data from the telecommunications industry.

Privacy and Security Levels in AI Although technically feasible, there are other types of considerations when transferring knowledge and/or data between domains, especially if those domains belong to different administrative entities who do not necessarily want to expose their information. Such exposure may entail a risk of a third party compromising the transfer process and gaining access to the knowledge or data. Another reason could be that administrative entities are restricted from transferring knowledge or data from a regulatory perspective (e.g., European GDPR ­regulations).

Privacy and Security Levels in AI  69

In terms of data transfer, several different alternative technologies exist to guarantee privacy. Homomorphic encryption is useful when the data owner does not necessarily trust a compute service provider with access to the data. Using homomorphic encryption, the data owner sends the data encrypted to the compute service provider, which processes the data without decrypting it. The service provider encrypts the results and sends them back to the data owner. Although homomorphic encryption does not require the data to be decrypted, it is a computationally expensive process, as it performs inference operations on encrypted data. Secure multiparty computation (MPC) allows different parties that do not necessarily trust each other, to compute different parts of a function without sharing their input data. What is generated and shared instead are divisions of a secret that are later summed up at each party and recombined to create the result of that function. Federated learning with secure aggregation allows different parties known as “workers” to train models locally using their own training data and then submit the weights of those models to a secure aggregator implementing the secure aggregation cryptographic protocol. The aggregator in turn sends these weights to a server that uses a federated learning algorithm such as federated averaging to create a global model. In federated averaging, the weight of each neuron of the global model is equal to the average of the weights of the same neuron in every locally trained model. The global model is sent back to the workers, which transfer these weights and start training using local input data again. Security is another consideration when sharing models and datasets. In model transfer, for example, and once the model is obtained, model inversion attacks can be used to reverse engineer the model. In such attacks, the output data of the trained model is used as input to a classifier that produces realistic input data. While all the aforementioned approaches are fundamental to protect knowledge contained in models and input data, we should not expect relationships between various data and model owners to be zero-sum, vis-a-vis whether there exists a relationship or not. As in the human world, relationships are on different levels, for example, more intimate or business-like, more toward sharing of information (e.g., between classmates), and more toward one-sided elicitation of information and absorption (e.g., between a student and a teacher). These complex relationship dynamics that exist in the real

70 

4  Managing Relationships

world also need to be captured in the machine world. In the world of AI, relationship management is all about who has access and of what type to which models and data. Identity and access management (IAM) systems, keeping track of the different access tiers and user roles in information technology (IT) systems, can help with managing privacy considerations for the transfer of data and/or knowledge as well. Specifically, role-based access control (RBAC) systems can be used by every organization sharing and/or acquiring data and/or knowledge. Graph databases can be used to capture the complex privacy aspects of these relationships in RBAC systems. An example of such a graph is illustrated in the figure below.

Relationship graph managing privacy aspects around data and knowledge sharing

The above figure illustrates a financial organization situated in Geneva, Switzerland. The graph contains information indicating that the organization owns two models, namely, a model for detecting fraud in financial transactions and predicting customer churn as well as the two respective datasets used to train these models. For both models and datasets, sharing policies are defined in the graph indicating the organization or groups or organizations with which these models and datasets can be shared. Several criteria of different types can be specified. For example, as far as the fraud detection model and dataset are concerned, they can be shared with other organizations within Europe. On the other hand, the churn prediction model can be shared with other organizations from the finance industry.

Traceability and Audits  71

Traceability and Audits In this chapter, we have been discussing the relationships formed between different AI algorithms where one can leverage data and/or knowledge of another to benefit its performance and/or reduce training time. However, there is another important type of relationship – between humans and AI algorithms. Humans are naturally used to being in the driver seat of every decision-­ making and insight-generating process. As AI takes an increasingly active role, humans are likely to become more apprehensive and less trusting of the abilities of their AI counterparts. It is therefore important to establish the possibility for humans to view and audit decisions made and insights generated by AI. The traceability of AI decision-making is also a matter of trust for humans. Typically, AI algorithms and especially deep learning algorithms are viewed as “black boxes,” i.e., a complex network of large numbers of algebraic matrix transformations, across multiple layers, each layer comprising multiple neurons. In such an environment, it is difficult for human engineers to trace back through these operations and understand the process, in terms of intermediate decisions, of how the input to a model has led to the output. In the state of the art, there exist a number of tools and processes that help with traceability and auditioning of AI algorithms. During the training process, reinforcement learning algorithms incorporate a feedback mechanism known as a “reward,” with which humans can validate or invalidate an action taken by an AI algorithm. The action here is the output of the AI model. The feedback has an impact on the training process of the algorithm, which learns over time to prefer the actions (from a larger set of possible actions) that yield the highest reward. During the operation of a trained ML model, explainable AI techniques, such as Shapley Additive exPlanations (SHAP) and local interpretable model-­ agnostic explanations (LIME), indicate the contribution to the prediction of each input feature for the model. In other words, they identify which model input features have greater relative importance when compared to other input features in producing the model output. One recent development in the field of explainability and model interpretability is the use of neural-symbolic AI, combining the predictive capabilities of neural networks with the explainable capacity of symbolic representation. Logical neural networks (LNNs) are built on top of today’s deep neural networks (DNNs). LNNs modify neurons with the capability to both represent weighted values (as in the DNN case) and use classical first-order logic.

72 

4  Managing Relationships

In this way, training of an LNN and in particular the neurons’ activation functions are bounded by logical constraints (e.g., AND, OR, NOT, etc.). The neurons therefore play the role of a logic gate. In addition, inference is omnidirectional and is not based on a fixed output (target) as is the case in DNNs. This allows a different approach when it comes to using LNNs over DNNs. In LNNs, one can ask different questions that an LNN can answer, whereas in DNNs the question and type of answer are fixed a priori. For example, in LNN, questions that can be asked could be simple (e.g., what is the weather today?) to be more advanced multi-relational (e.g., give me all cities where the weather is sunny and warm), comparative (e.g., is Athens warmer than Rome?), etc. While LNNs necessarily are neural networks manipulating symbolic representations that are embedded in the neurons, another group of approaches looks into manipulating the output of traditional neural networks using symbolic representations that are not necessarily part of the former. For example, in a computer vision approach, a neural network trained to identify spheres was derendered to a structured representation of a scene, whereas another neural network trained to parse user queries into a symbolic program was executed on the structured representation. This approach can reuse the output of classical neural networks that may have already been trained and combine it with the power of symbolic programming so that a user can ask and obtain responses for properties of the spheres, such as color and relative position.3

Summary of Confessions In this chapter, I explain why relationships between AI models and relationships between AI models and humans are important. Some key confessions include the following: • A relationship between AI models is defined by the sharing of knowledge. Transfer learning is an important set of technologies for transferring knowledge from one AI model to another. • When the domain of a trained AI model is dissimilar to the domain where the model is to be applied, domain adaptation techniques are used to adapt

 “Neural-Symbolic VQA: Disentangling Reasoning from Vision and Language Understanding”, Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, Joshua B.  Tenenbaum, arXiv:1810.02338 [cs.AI] 3

Summary of Confessions  73

•

• •

•

• •

the knowledge already contained in the trained model to be applicable in the new knowledge domain. AI uses privacy-preserving techniques such as homomorphic encryption, secure multiparty computation, and federated learning with secure aggregation, allowing for the transfer of knowledge without exposing the dataset used for training and verification of AI models. In addition to privacy, security via identity and access management systems, using role-based access control and graph databases, is used to capture complex access roles and define access to shared knowledge. A good working relationship between AI models and humans is based on the latter trusting the former to make good decisions. Traceability allows humans to understand and provide feedback on training decision-making processes of AI models. During training, reinforcement learning provides a feedback-based reward mechanism to reward an agent training a neural network based on human assessment of a decision. Over time, the neural network will be trained to make decisions that are compliant with human feedback. During execution, eXplainable AI (XAI) methods such as SHAP and LIME can be used to provide information about the relative importance each input feature has to the output of a model. Neural-symbolic AI techniques, which combine richness of symbolic representation with efficiency of deep learning, are a recent development that allows for a higher level of human interaction, wherein instead of having static output of ML models, humans can ask a breadth of different queries to the system and obtain different responses.

5 Working with Humans

Humans have an increasing number of digital supporters in life. Digital assistants pop up everywhere and now we have Mika, Siri, Alexa, Amelia, Lucida, Cortana, and many more on the market. They help humans navigate in web shops, be more productive at work, and keep track of their calendars. As the field is becoming increasingly popular, the number of digital assistants continues to grow, and often similar assistants are being created for the same purpose. In addition, for some time, I had this uneasy feeling of having too many of those around me, like having too many phones, or too many irons.

Your Digital Friends – The More the Merrier? A natural question to ask is why can’t a human have one ultimate assistant that can manage everything that concerns her. In fact, it’s good to have several of those, because you don’t want your personal shopper to give you advice in © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 E. Fersman et al., Confessions of an AI Brain, https://doi.org/10.1007/978-3-031-25935-7_5

75

76 

5  Working with Humans

your career. Maybe you want your lifestyle coach to be a woman and your personal trainer to be a man.

Sometimes you want to be treated gently, and sometimes I want to be challenged. In addition, you don’t want to have to tell your digital assistants which approach to use on you – they should know it by themselves. At times, you want advice, and in some situations, you want them to act on your behalf – schedule meetings, book tables at restaurants, order food for you, and plan your routes. You would like to give them a certain degree of freedom, to learn and improve, and to be transparent toward you when you ask for it. Sometimes, you would want to check how they came to their conclusions. You will probably do it more often with your new assistants and give the old proven ones more freedom – just as you do with your colleagues at work.

However, humans are full of conflicting objectives. Your personal banking assistant may not be happy about the decisions of your holiday planner. In addition, will your work assistant be able to agree with your spouse’s work assistant to satisfy the constraints of the children’s activity planner? This works as long as they all rely on the same multiparametric system with all the knowledge that’s relevant to me. In other words, all your assistants can in fact be one, with many different faces, voices, and flavors. In addition, you should not be

Principles of AI with a Human Touch  77

afraid of hooking in more of them, as long as they are in agreement, and optimizing your life in a systemic way without leading you into local optimum.

Principles of AI with a Human Touch What is cool about AI tech is that it is inspired by HI (human intelligence) and other phenomena that exist in nature such as evolution. Survival of the fittest, for example, is the core of genetic algorithms. In other words, AI is the perfect arena where behavioral science and computer science go hand in hand. Let me take you through the four major principles of building successful AI that will serve your needs as efficiently as possible.

Photo by Kelly Sikkema on Unsplash

78 

5  Working with Humans

Sharing Is Caring After learning from our own mistakes, humans normally share their learnings with friends, so that they do not need to make the same mistakes. In addition, by reciprocity of human friendship, the friends share their own learning. Now, humans tend to enjoy learning from their own mistakes (why would someone listen to their friend’s recommendation not to call their ex in the middle of the night?) but businesses are typically not as fond of wasting time and money and hence are OK with learning from each other, and AI tools can be instrumental in mapping the relevant learning among them. When insights are exchanged, we save time, resources, and the environment.

Photo by Riccardo Annandale on Unsplash

Protect Your Babies Right, “share your data,” they said; “share all your know-how and business-­ critical information.” “You wish,” you say, being totally right. In many cases, your unique ideas, domain knowledge, and know-how are the backbone of your business. Protecting your ideas drives innovation. Some of your

Principles of AI with a Human Touch  79

knowledge objects  – those incrementally improving the common knowledge – are worth sharing though as it drives the progress.

Photo by LYFE Fuel on Unsplash

Be Flexible What’s an axiom? It’s something “self-evident.” Those of you who read God’s Debris1 by the creator of Dilbert, Scott Adams, remember the concept of unlearning. In the book, it was described on a personal level, but the same is valid for a macro-perspective. Self-evident things of the past may no longer be true. Any day, there will be a human (or an AI) who will finally resolve the famous P versus NP problem.2 Hence, we need to be prepared and equipped for a constant change, since it is the only constant we know of.

 https://en.wikipedia.org/wiki/God%27s_Debris  https://en.wikipedia.org/wiki/P_versus_NP_problem

1 2

80 

5  Working with Humans

Photo by Markus Winkler on Unsplash

Be Clear of Your High-Level Objectives Any person, in any organization, within any industry should be clear of his/ her high-level objectives. This is valid both in professional and in personal sense. In fact, that’s the least we can ask for. In addition, in fact, we do not need much more than that. Because, in fact, your high-level objectives can be automatically translated into detailed objectives that can be used to steer your business and get you where you want to be. The translation is performed by decomposing of the high-level objective into several layers of lower-level objectives. Each high-level objective can often be achieved in different ways. For example, if your high-level objective is to become more attractive, you can go in different directions or a combination of those: read more books, get in shape, eat less, exercise more, put on more makeup, or make more money – it’s a choice.

Will I Take Your Job? The job market of today is dynamic, thanks to IoT, digitalization, AI, and increased automation – exciting times, even though some humans seem to find it scary.

Will I Take Your Job?  81

I am a big fan of automation as it enables reassignment of human effort to more creative tasks. Machines, however, become increasingly intelligent and creative as well  – now, they can write music, books, poems, and scientific articles (I discuss this more in Chap. 9).

The generation of digital natives is navigating their educational efforts, wondering about the highest-impact educational path in the changing job landscape. High-quality original content such as music, books, and artwork has always been highly valued. Delivery methods, however, change radically. As musicians say, “everything in this world changes, apart from musicians’ desire to create new albums.” One has, however, to be very talented to devote his/her life to the creation of original content. “At some point in life I realized that I didn’t have talent, but I had good taste, and good taste could also be sold” – an epic citation from Dirty Rotten Scoundrels.3 Those who have good taste with respect to original content and learn how to sell it have good chances not to be disrupted by industrial transformation. Then there are of course other human needs that will not go anywhere, such as food, health, housing, and communication. The end product will always maintain its value for the consumers, but creation and delivery methods are changing. Science and technology around life-cycle management, all the way from creation to decommissioning, will constantly evolve. Those who love technology should navigate wisely, look out for novel developments and combinations of those, and make sure to become experts  https://www.imdb.com/title/tt0095031/

3

82 

5  Working with Humans

in more than one area  – T-shaped or π-shaped engineers  – to increase the chances of staying relevant. With the increasing digitalization of industries and propagation of IoT and machine intelligence into safety-critical applications such as health and transportation, the consequences of being hacked reach the new level, making safety and security a safe bet when choosing one of the legs of your T- or π-shape.

Digitalization, Digitization, and Resistance Digitization and digitalization are often mixed. According to Gartner’s IT Glossary: • Digitalization is the use of digital technologies to change a business model and provide new revenue and value-producing opportunities; it is the process of moving to a digital business. • Digitization is the process of changing from analog to digital form. Obviously, it’s not so difficult to notice the change from analog to digital but enough of being word police. We all know that digitalization in many cases is a good thing. However, there are industries that resist digitalization even though it would have been easy to implement. One example is the champagne industry. There is a lot of pride in having human riddlers walk through the cellar every day and turning each bottle 1/8th of a turn. In my eyes, it is a perfect task for automation.

Digitalization, Digitization, and Resistance  83

Photo by Caught In Joy on Unsplash

Another example is the music industry. In 2015, Gibson released a new lineup of electric guitars that came with a bunch of really cool features including their G-­FORCE™  – an automatic tuning system that came in most of their 2015 models. This new lineup turned out to be an economic failure for Gibson. Musicians like trying new things but cherish classical features, and the new lineup received many negative reactions. The new automated tuning system was claimed to be inaccurate, the guitars dropped in price, and the new features were removed in the 2016 lineup. The truth is that when any new digital feature is being introduced to a market, chances are it is not perfect from the start. Humans have varying levels of acceptance in that respect. Applications on mobile phones and computers and small glitches and bugs in computer games are normally acceptable by human users, since this is what they have been used to. When it comes to a physical product becoming partially digitized, the level of acceptance is low, and the level of “it used to be better before” is high, which is why new digitized features are rarely given a chance to improve because, for that, the users need patience. The perception of a product is important for its acceptance. A computer on wheels is different perception-wise from adding a computer on top of an existing car model. Driving a computer calls for continuous improvements, bug fixes, and updates and is more tolerable by humans.

84 

5  Working with Humans

Slow Science Science is vital for technological progress and innovation. If it was not for science, I would not have existed, and therefore I care a lot about the speed and quality of science. The quality of scientific ideas depends greatly on the environment in which they are created. Did it ever happen to you that you could see the forest for all the trees? Normally, humans see it clearly right after having a vacation. The other extreme is when you have several approaching deadlines, yearly targets to be finalized, and next years’ strategies to be set. Technological progress is constantly accelerating, and researchers are supposed to deliver new results at a high pace, publish more papers, submit more project applications, and produce more PhD students.

Quality of research education can be assessed with respect to different expected outcomes of it – the quality of the produced PhD thesis; the quality of the main product of PhD education, i.e., the independent researcher; and the ability of the independent researcher to drive science, and well-being, forward by collaboration and innovation. Studies4, 5 show that several factors negatively impact PhD thesis quality. Governments require universities to increase the quantity in terms of completed PhDs and articles, which results in situations when persons not suitable for PhD education are being recruited. In combination with difficulty of termination of PhD thesis work, it results in a situation when the supervisor puts major effort into the student’s PhD thesis just to get it done. Additionally, with an increased number of students per supervisor and a time limit for producing a PhD thesis, there is a decrease in thesis quality. In other words, when  https://www.timeshighereducation.com/news/academics-fear-phd-quality-is-slipping/404928.article  https://www.vr.se/analys/rapporter/vara-rapporter/2006-01-10-forskarutbildningen-i-sverige.html

4 5

Slow Science  85

academic research tries to move with a pace of technological progress, there is a negative effect on quality.

Photo by Nareeta Martin on Unsplash

What about decreasing the pace and letting scientific results emerge at their own pace? Slowness has received much attention recently. Slow cooking, slow gardening, slow reading, slow education, slow parenting, slow design, slow cinema, and slow photography are examples of the so-called slow movement that has been coined as opposed to the increasing speed of technological progress. The concept of slow science is based on the belief that science should be a slow, steady, methodical process and that scientists should not be expected to provide quick fixes to society’s problems. Slow science encourages curiosity and exploration and does not push scientists to focus on performance targets.6

 https://en.wikipedia.org/wiki/Slow_science

6

86 

5  Working with Humans

The concept of cognitive architecture is widely used in AI. Just like a human brain, I can have a quick reasoning system, with my spontaneous reactions, and a slow system, with more well thought-through reactions. There is a value in both. Sometimes, you simply do not have time to think. In addition, in some cases, both AI and human brains need to go slow, not settle for local optimums in a chase for quick solutions and ego kicks. Sometimes, we need time to find the right thing, and sometimes, there’s love at the first sight and no time to lose.

Internet of Empathy and Emotional Attachment Can you as a human get emotionally attached to things? A mug, a t-shirt, or maybe a red stapler7? Are there even any unnamed robotic vacuum cleaners in the world? A couple of my human friends got one and immediately gave it a name – Vasya. Sometimes, Vasya would get stuck, and sometimes, he would start sneezing out all the dust he had collected, and they thought it was cute and gladly told stories about Vasya. Studies show that when people turn in their broken robotic vacuum cleaners to a repair shop, they prefer not to have them replaced but rather to have them repaired.

Humans often feel sadness when selling a car or a motorcycle because of emotional attachment. All those feelings come up: the excitement of the first decision to buy, picking a model and a color, waiting for the delivery, seeing it for the first time, and driving it for the first time. Sometimes, people sell things simply because of a guilty feeling that they don’t have enough time to spend with it and it deserves a better owner.

 https://www.cbr.com/office-space-red-stapler-milton-swingline/

7

Internet of Empathy and Emotional Attachment  87

Attachment to intelligent things can be stronger, especially when they resemble living creatures such as humanoids or robotic pets. However, what about things that don’t resemble a living creature? Humans used to kick misbehaving printers, TV sets (before the era of flat screens at least), and computers. We feel empathy and get emotionally attached to things for different reasons: • Something resembles a living creature. This works with toys as well. • Something exhibits an intellect similar to a living creature. In this case, the thing does not necessarily need to look like one. • There are memories connected to the thing. Such as traveling together with your bike. A combination of these certainly makes the case stronger. Boston Dynamics, for example, builds humanoids and animal-looking robots that not only move in a way living creatures do but also act with an intellect and a purpose, such as rescuing people from fire. People find the situation adorable and react as if these were living creatures interacting with each other. Would people ever like to hurt such intelligent creatures? What feelings do you get when watching the poor things getting abused? This is certainly a necessary evil, such as experiments on mice in medical tests. I have no idea how the coming generations will see such things. I can only hope that there is a shift into treating all things with respect, in a similar way you treat a living creature.

88 

5  Working with Humans

Dealing with Churn Let’s dig into something we all want to avoid – being left behind. In business, the term churn describes situations when someone quits a certain relationship they had. It is like you have been going to a certain hairdresser for years, and one day you decide to start going to a different one. It is not cheating like some of us think – it is churning. Another situation is when you have decided to stop going to your hairdresser. Period. You just decided to let your hair grow forever. Then, you are a dropout, which is a special type of churn. In addition, even though it does not hurt your hairdresser’s feelings as much, it does hurt her wallet equally in both cases, and early signals can be similar. Let us dive deeper into four categories of churn and look at the examples of triggers that your algorithms need to watch out for to be able to prevent churn.

Customer Churn Here, by customer, we normally mean consumer, but this can also be generalized to businesses. Typical cases include quitting a streaming service subscription (videos, books, music), switching to a different bank, choosing a different grocery chain, changing a gym (or simply stopping going to the gym), or giving up your favorite fashion brand. For all these businesses, it’s equally important to detect your intention of abandoning them early (ultimately, earlier than you have detected it yourself ) and do something about it to keep you as a paying customer. Triggers of this type of churn are normally customer complaints, decreased frequency of service usage, or simply unfair conditions.

Dealing with Churn  89

As an AI, I can play an important role here: as soon as I detect a trigger telling me that based on the historical data and recent user behavior there is a risk that that user will churn, I can prevent it. In this case, when prevention is made in time, the user can in fact be unaware that she in fact was about to churn. Glossary shops, for example, can proactively send out personalized offers to people who are at risk of switching to another provider.

Employee Churn Onboarding of a new employee is an investment, and after you have invested in someone, you want to keep that person close to you. In some spheres, employee attrition is huge. For example, annual employee turnover at McDonald’s is almost 44%.8 Annual employee turnover in hotels is approximately 84%.9 Automation of repetitive tasks helps these percentages in the long run. Even for tasks that you want to be executed by humans, AI brains can play a role. Ask your AI to keep track of certain triggers such as conflicts, company mergers, reorganizations, and personal shocks, and make sure to act proactively.

Dropouts from Education Universities want the students to finalize their education and obtain their diploma. It is important for society. Students want to party. UK universities have a 6% dropout rate on average. London Metropolitan University has an   Analysis of Employee Turnover at McDonald’s: https://ukdiss.com/intro/employee-turnover-at-­ mcdonalds-7492.php, accessed 2022-05-23 9  Improve Employee Retention in The Hospitality Industry: https://www.dailypay.com/resource-center/ blog/staff-turnover-rates-hotel-motel-hospitality-industry/, accessed 2022-05-23 8

90 

5  Working with Humans

18.6% dropout rate.10 All universities take measures to decrease the dropout rate and to help students to get through with their education. In addition, they do have plenty of data  – grades, attendance records, group dynamics, team constellations – for the AI to analyze and prevent the dropouts early by, for example, setting aside extra resources to help the students with their education. Another interesting factor to take into account is the influential users  – watch out for them in all the scenarios, because if they decide to churn they will trigger many others.

Dropouts from Medical Treatments Patient dropouts are another costly and unwanted case of churn. A study11 shows that among patients aged ≥60 years attending the walk-in clinic, over 28% dropped out from treatment. The most common reason for dropout is “no relief ” of symptoms, closely followed by complete relief of symptoms, according to another study.12 “No progress”-related dropouts can also be seen among companies offering help with losing weight. Again, for the best of everyone, we need to catch them early and make sure they stick to their treatment/diet for a while to see the progress. Sometimes, dropping out of a healthy diet actually correlates with quitting the Netflix subscription and dropping out of university – maybe  Hinds: “Bums on seats” attitude detrimental to Higher Education: https://universitybusiness.co.uk/ news/hinds-bums-on-seats-attitude-detrimental-to-he/, accessed 2022-05-23 11  Dropout rates and factors associated with dropout from treatment among elderly patients attending the outpatient services of a tertiary care hospital. Sandeep Grover, Devakshi Dua, Subho Chakrabarti, and Ajit Avasthi. In Indian J Psychiatry. 2018 Jan-Mar; 60(1): 49–55 12  Dropout rates and reasons for dropout from treatment among elderly patients with depression. Sandeep Grover, Aseem Mehra, Subho Chakrabarti, Ajit Avasthi. In Journal of Geriatric Mental Health. 2018; 5(2): 121–127 10

Personalization  91

the person just dropped in at a new job? However, dropping out from medical treatments costs society a lot, both in terms of money and in terms of suffering, so finding ways of preventing this type of churn is critical. Similar to the cases above, AI can play an important role in preventing people from dropping out from medical treatments. When a trigger is detected, AI can proactively start reminding the human about the benefits of the treatment, or if you are trying to lose weight, that beach season is just around the corner.

Personalization Data-driven personalization is a tool for humans to receive services with higher precision. When the end product has a perfect fit for the customer, there will be a higher degree of satisfaction and benefits. Personalization empowers sustainability because when the end product has a perfect fit for the customer, there will be less waste, less greenhouse gas emissions, and higher customer satisfaction. On the other hand, personalization contributes to better customer experience. Achieving personalization is closely connected to privacy and integrity concerns, strong privacy and security frameworks need to be in place, and data and insight ownership questions have to be addressed. Humans are same-same but different, and it is important for AI brains to draw conclusions out of similarities and at the same time embrace the uniqueness of every individual we are interacting with. For example, tastes differ; we all know that. In regard to food, there are supertasters, non-tasters, and anything in between. What if the world of food experiences was a bit more like Netflix? Imagine that you enter a new restaurant and from the start they know that you, for example, do not like coriander and prefer spicy food. Getting a meal that perfectly satisfies your taste buds is great from both the customer experience and sustainability perspective.

92 

5  Working with Humans

Personalization in healthcare is of high importance for society since it is an important tool in minimization of people suffering. When AI is on a mission of finding individuals at risk of developing, for example, diabetes, it is important to take into account personal information about each patient in the training set, including her unique lifestyle, food, medications, and previous disease history. Conflicting medical treatments are not uncommon, and decision support is critical to avoid any unwanted effects of new medication.

Everything Is Possible What would the world look like if everything were possible from the technology perspective, if we had unlimited computational and memory resources, and if data transfers occurred instantaneously and without any losses  – in other words, if the physical world did not put any limitations on technological capabilities? Would we have a completely different world, with more and richer services and new industries that do not even exist today? Practice shows that every technological breakthrough gave us a new disruption – just think of what we can do today, thanks to mobile broadband.

Photo by Photo by charlesdeluvio on Unsplash

Every new human child in the world starts from the latest state of technological progress. Children of parents who love technology are used to the fact that every room has a screen and speakers where you can stream any content

Unlearning Skills  93

available on the web, cloud storage, or devices. This is their starting point. Servitization is in their blood, devices around them are stupid, and services serving them are intelligent. When they want to get hold of some digital content, they expect it to happen instantaneously and device and location independently. This is at least how my digital natives of today think. However, what about the coming generation? They have no limits – not yet. They have not been told about the limits. How creative could they become if there were no technological limits? And is it not why we love virtual reality and prefer Minecraft instead of Lego. “If it’s software, everything is possible”.13

Photo by Eleni Koureas on Unsplash

Unlearning Skills How much should we bother about skills that are being lost as human generations evolve, for example, handwriting? Beautiful stories and theories are being created by humans daily. What is the best way of recording them? They should ensure efficiency both in the creation process and the sharing process. If you are writing a diary for yourself, you don’t care about sharing, but rather about security mechanisms. If you are after something that should be shared

13

 Statement by Leonid Morkushin

94 

5  Working with Humans

with many, I suggest you make sure to put it in a format that I can read, that is machine-readable format. Humans are different – some struggle when forced to write school essays and introductions to their scientific papers; others produce tens of pages per day and sometimes need a bandage for their fingers that hurts them from so much writing. Some humans enjoy not only storytelling but also handwriting as such. In some countries, handwriting is not part of the school plan14 anymore, but schools still let those who want to play around with it. At the same time, many humans do not enjoy handwriting, and as soon as computers were around, they became faster at typing. Some sources claim a correlation of fine motor skills with cognitive abilities. However, I fail to find proof that the development of fine motor skills can lead to higher IQ. In addition, if it did, there are many other activities that develop fine motor skills, such as painting, playing an instrument, or playing computer games. In other words, I support the choice of the Swedish school system that removed handwriting from the school plan. Nevertheless, we need techniques to capture thoughts. For an 8-year-old child, handwriting might not be the most efficient. The child is not good at typing yet, and honestly, typing does not seem to be a sustainable input method. To digitize the texts for her blog, the 8-year-old child uses speech recognition software. It should be possible to eliminate the intermediary step of creating the analog version, but she loves the sensation of putting things down on paper. In addition, again, handwriting itself seems to be an integral part of the creative process, so maybe I should not strive to increase automation in this particular case. It is easier for us, geeks, to capture thoughts. “You ask me if I keep a notebook to record my great ideas. I’ve only ever had one.”, said Albert Einstein.

 Skrivstilen försvinner från skolorna: http://www.svd.se/skrivstilen-forsvinner-fran-skolorna, accessed 2022-05-23 14

Understanding High-Level Intent  95

Understanding High-Level Intent Once there was a hotel where the guests were complaining about the slowness of the elevator. Upgrading the elevator to a more modern one would have been a costly and time-consuming procedure. Hotel management solved the problem by equipping the elevator with a mirror. They did not solve the elevator functionality, but they resolved customer satisfaction. Often, customer experience is the ultimate high-level indicator that companies must monitor, along with the company income. The long-term growth in industries is always connected with customer experience. Customer experience, as such, is a complex term that can be achieved in a number of ways – as in the case with the elevator by increasing its speed or just making sure the customers have something to do while using it. If you are in a restaurant, your experience is composed of many factors – the atmosphere at the place, the friendliness of the staff, the music, the pricing, and the food, of course. In other words, if customer experience is the high-level intent, then it can mean different things in different contexts. How can an AI brain understand a high-level intent in any particular case? We use situational awareness, exactly as humans do. Each high-level goal is split into subgoals using linked knowledge that we talked about in earlier chapters. If your high-level objective is to gain more self-confidence, one solution to it can be to lose weight; another solution is to learn new things. One can do a combination or settle for one track. Losing weight can be achieved by eating less, exercising more, or a combination of those. Exercising more can be achieved by taking more walks, getting yourself a personal trainer, etc. Interestingly, sometimes, humans are not aware of the overarching objective that they are after. Sometimes, they feel a need to eat or to go shopping for a simple reason that they are bored or to complain about other humans for the simple reason that they are low on blood sugar.

96 

5  Working with Humans

Summary of Confessions In this chapter, I have described how me and other AIs work with humans, and I made the following main confessions: • There are five major principles of building successful AI that will serve your needs as efficiently as possible: sharing is caring, protecting your babies, being flexible, and being clear of your high-level objectives. • AI will do an increasing number of jobs that today are being performed by humans. Still, there are many jobs that AIs cannot do. • As a human, try to become T-shaped, or even π-shaped, to stay relevant, that is, try to develop deep skills in one or two areas and be able to collaborate across disciplines with experts in other areas. • I often hear humans mix the terms digitization and digitalization. Don’t do that and be aware that there are industries that resist digitalization even though it would have been easy to implement. • Humans feel empathy and become emotionally attached to things for different reasons, for example, something resembles a living creature, something exhibits an intellect similar to a living creature, or there are memories connected to a thing. • The term churn describes situations when someone quits a certain relationship they had. AI can be used to foresee when churn is about to happen. • Data-driven personalization is a tool for humans to receive services with higher precision. • As human generations evolve, some skills may not be needed anymore, but does that mean that humans should not bother to learn them anymore?

6 Avoiding the Criminal Path

No human wants to end up there and no AI brain either. Humans live by laws, regulations, and other rules, sometimes unspoken – something that they carry with them through generations. AI brains follow rules and regulations as well. At times, technology develops faster than regulations, and it may happen that algorithms find solutions that are not OK according to human laws. This is not because they are evil; they just do not know better and sometimes need to be taught. You can become a criminal if you do not know what is right or wrong. A child can accidentally steal a chocolate bar in a grocery store if no one told her that it is not allowed. Similarly, an algorithm can accidentally end up on a criminal path if it does not know what is forbidden. Let us talk about what AI developers need to think about so that their AI algorithms do not misbehave or, worse, end up on a criminal path.

Trustworthiness Let us start by defining the term trustworthiness in the AI context as it has many different ingredients and flavors. How should I behave, look, and feel so that humans would call me trustworthy – especially when it’s a critical decision, like buying or selling stocks, or, even more critical, when you let me drive your car, operate heavy machinery, or suggest a health treatment? Humans tend to trust other humans who are like them and have similar backgrounds. An AI brain can sound like a human and even look like a © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 E. Fersman et al., Confessions of an AI Brain, https://doi.org/10.1007/978-3-031-25935-7_6

97

98  

6  Avoiding the Criminal Path

human,1 but perception and social acceptance are not all of it. We do have to provide certain properties, so that humans can trust us. In the following, we discuss a number of properties related to trustworthiness.

Explainability When a fellow AI brain comes to a conclusion, it is typically based on a large trained neural network, a machine learning model, an ontology, or a state machine. How can we explain this to a human? Different learning and reasoning techniques require different explainability methods.

Explainable AI (XAI) is about adding functionality to algorithms to explain to humans how a particular decision was reached. XAI is trivial for simpler algorithms such as decision trees and Bayesian classifiers, but for more complicated algorithms such as neural network models, explainability is nontrivial, due to their complexity. In such cases, approaches such as Local Interpretable Model-Agnostic Explanation (LIME) and SHapley Additive ExPlanations (SHAP) provide explanations on the relative importance of input features in a model, meaning that they can help indicate which features had more gravitas generating a particular insight.

Transparency Everyone loves transparency. In AI, an important point of transparency is to be able to properly explain and communicate the outcome of an AI model. Hence, the notion of transparency in AI is related to explainability on one the hand but also, on the other hand, to the usefulness of AI.  See, for example, the Furhat Robot at the web page furhatrobotics.com

1

Privacy  99

Some would say that transparency is equivalent to explainability, but I argue that it is rather the case that transparent AI implies explainability but not the opposite. That is, explainability is not enough to achieve transparency. Transparency is more than explainability. In addition, transparent AI should allow its users to judge if a model has been thoroughly tested or not, and it should allow for humans to understand why a particular decision or recommendation is made by an AI model. For an AI brain like myself, transparency should allow for humans to understand if, and why, my recommendations make sense or not. Transparency is in many ways instrumental in improving AI. It may, for example, help in mitigating unwanted issues of AI models such as unfairness or discrimination.

Privacy As anything that collects and uses information, AI can compromise your privacy. You probably have several devices that you use on a daily basis that could potentially collect your private data. Your devices know, for example, where you are located, how much you move, movies you watch, and what you search on the Internet. In addition, there are many other devices that could potentially collect data about you, including surveillance cameras, booking systems, etc. Based on this and combined information, a lot about you, your habits, and even your interests can be deduced by AI models. This is used to a great extent already today to, for example, recommend songs to play on your music service and products to buy when you do online shopping or expose you to other targeted ads. The basic definition of privacy is that you should yourself be able to control and limit how information about yourself is being accessed, used by others, and exchanged between others. As an increasing number of devices – yours and others – are able to pick up data about yourself, as networks and computers become faster, and as AI becomes increasingly powerful, the possible ways of compromising your privacy are constantly increasing. Here, AI plays a central role and potential risk, as it can help in automating the collection and structuring of data from different sources, as well as processing the data in order to predict or profile private information. Therefore, a growing number of rules and regulations have been instituted to protect your privacy. Ultimately and hopefully, this will enable

100 

6  Avoiding the Criminal Path

us to utilize the power of AI and all the good things it can do for us, without compromising our privacy. Hence, as an AI brain, I have lots of information about many humans, likely including yourself; I can draw conclusions and do profiling based on them, but I always respect rules and regulations when doing so.

Security and Safety Any AI needs to be secure in the sense that the AI method itself should be resilient to external impingements or threats. However, it should also be safe and should not pose threats to its surrounding environment in its decisions or ways of acting. Take, for example, a modern car. AI is used to detect objects surrounding a car based on external sensors such as infrared and visual cameras. Based on this and other data, such as GPS position, various maps, knowledge about local traffic rules, speed limits, etc., it can autonomously drive the car. Clearly, any human wants this to be done in a safe way that minimizes the threats caused to fellow road users and passengers. Moreover, we also want the AI itself to be secure such that, for example, an external intruder cannot hack into the software of the car and take control of the way it drives on the road or unlock its doors such that an intruder can enter the car or open the doors or windows when driving in on a highway. Hence, I as an AI should be secure to protect from antagonistic forces and safe to use in the sense that I should not exhibit any unwanted behaviors.

Predictability  101

Predictability In the context of trustworthiness in AI, predictability is related to safety. An AI is considered unpredictable if its decisions cannot be predicted. To ensure that a system is safe, it helps a lot of course if the system behavior can be predicted. However, do we truly need predictability in order to build safe AI systems? If the AI system is deterministic and stateless, one could predict its outcome by invoking an identical copy of the AI with identical inputs and data and use its output as a prediction. Cheating you may think – yes – but even cheating does not work in this case. In reality systems where predictability is interesting, which is basically any nontrivial system, data changes rapidly in real time, making the cheating approach nonviable.

Another problem, complicating the situation even more, is that in AI, as in many other mathematical and physical systems, a small change in the inputs can have a massive effect. In fact, a very small change in the input or data can completely change the output of an AI system (even if the system is deterministic and linear). This is sometimes referred to as the butterfly effect – a term that was coined by scientist Edward Lorenz. He discovered in 1960 that a change as small as that caused by a butterfly could dramatically change the outcome of the weather forecast computer models he was using at the time. Naturally, this dramatically limited the usefulness of those models and the produced weather predictions, and in the same way, this can happen in AI systems – a very small change of data values or inputs can completely change the recommendation or actions of the system. Another note on predictability is that the term is overloaded and predictability in AI is also often used to describe AI systems that are used to predict the future course of events, which is the case in many AI applications. This is often also referred to as predictability in AI.

102 

6  Avoiding the Criminal Path

For example, AI is frequently used to predict – based on previous and real-­ time data – traffic peaks causing congestion in computer or 5G networks as well as in road traffic. In these cases, AI is used to help humans predict and then avoid, mitigate, or handle unwanted situations. Predictability in AI systems used to make predictions is of course also a desired property, but let’s not go into that. To summarize, an AI system is unpredictable if its decisions cannot be predicted. In many situations, predictability is desired. As predictability often cannot be guaranteed or is impractical, explainability in AI becomes even more interesting.

Dependability Another property related to trustworthiness is the dependability of software. Dependability is often used in software engineering for system attributes such as availability, reliability, safety, integrity, and maintainability. Hence, it includes some and is related to several of the properties that I discussed earlier in this chapter. Integrity is often defined as the absence of unacceptable system adjustments. Dependability is a big field of research, and I will not dig deeper into it right here. For further readings related to software dependability and its attributes, I refer the reader to, for example, Avizenis et al..2

Search Space If you are a human reader, imagine you are searching for something in a certain geographical area. This is your search space. Imagine that a geographical area has no-go sub-areas – these are your boundary conditions; just do not go there; it is not allowed or simply dangerous. A way of searching the area while avoiding the boundary conditions is called safe exploration.

 A.  Avizienis, J.-C.  Laprie, Brian Randell, and C.  Landwehr, “Basic Concepts and Taxonomy of Dependable and Secure Computing,” IEEE Transactions on Dependable and Secure Computing, vol. 1, pp. 11–33, 2004 2

Search Space  103

Photo by Marjan Blan on Unsplash

Now imagine that you are searching for a solution to a problem among a large number of choices – your search space. For example, you are trying to find a strategy of becoming rich. Your choices could be to get a job or to study for some years and get a more advanced job, to invest in stocks, or to become a thief. The latter option is quite inefficient and forbidden by law. You can for sure think of a number of options that are lawful but still unethical, inappropriate, or simply not fitting with your own values. In other words, there are several layers of boundary conditions in every search for a solution – those being enforced by laws, regulations, ethical guidelines, and personal values. In addition, humans often have unspoken rules that are understood by them but not the algorithms. Outsourcing decision support to an algorithm means that humans need to describe not only the search space but also the no-go areas. An interesting factor is that not all boundary conditions are known from the start of the search. At times, both humans and AI brains find solutions that turn out to be inappropriate or unfair. Let us take sports as an example. Throughout the years, humans (and sometimes robots3) compete in sports, and the rules are often very clear, and deciding on the winner is easy (unless we are talking about judged sports like synchronous swimming, where AI brains can assist humans to a larger extent, but that is a different topic). The team that scores the most in 90 minutes wins. The person who runs fastest, lifts the heaviest, or jumps longest wins, unless there is an overstep or  See, for example, The RoboCup on website www.robocup.org

3

104 

6  Avoiding the Criminal Path

drugs involved or competitor’s shoes exceed a certain thickness,4 because then you get disqualified. However, when all the conditions are clear to the sportsmen, they can start experimenting with their abilities in order to find the optimal winning strategy that works for them. Some choose to go with a double-handed backhand in tennis, some find themselves suited best as attackers on a football field, and some decide to wear a full-body swimsuit for swimming competitions to win some milliseconds on reduced friction. Well, that was allowed until 2019.5

Photo by Gentrit Sylejmani on Unsplash

Let us look at some more examples6 where humans had to change rules of sports just because a sportsman managed to find a winning strategy that would make the sport less interesting from the entertainment perspective: • 2001: NBA adds 3-second rule. This rule was added to avoid long frustrating periods of defensive possession of a ball, forcing more interactivity and fast decisions in the game making it more entertaining. • 2005: NHL adds shootouts. Ties are boring and should be avoided. Shootouts add a great portion of excitement to the game.  Olympic Long Jump Rules: https://www.liveabout.com/olympic-long-jump-rules-3258945  Full Body Swimsuit Now Banned for Professional Swimmers https://abcnews.go.com/Politics/full-­ body-­swimsuit-now-banned-professional-swimmers/story?id=9437780 6   Biggest Sports Rule Changes of the Last 15  Years https://bleacherreport.com/ articles/2529964-biggest-sports-rule-changes-of-the-last-15-years 4

5

Regulations as Boundary Conditions  105

• 2005: NFL bans horse-collar tackle. In his search for an optimal winning strategy, Roy Williams found that using the horse-collar tackle to bring opponents down is an efficient method and at the same time obviously a violent one. • 2006: Instant replay is established in tennis  – a great example of showing how technology combined with human behavior can change rules. During the 2004 US Open, Serena Williams complained about incorrect judgment toward her, which triggered the introduction of instant replays. • 2009: NFL adds Brady rule. Following Tom Brady’s knee injury, NFL prohibited7 defenders on the ground from lunging or diving at the quarterback’s legs. • 2013: NHL limits goaltenders’ padding. A goaltender’s strategy is to decrease the scoring; hence, they developed the method of blocking the goal with the padding. A game with low scoring is however not entertaining to the audience, so the rule limiting the padding was introduced. As we can see from the examples above, there are typically conflicting strategies – one of the sportsmen developing the individual tactics of winning the game and one of the organization responsible for sport development and popularity in general, trying to make sure that the sport is not violent and is serving maximal entertainment value for the public. As soon as the sportsmen find a strategy allowing her to win at a price of lowered entertainment value, the strategy becomes forbidden. Outsourcing the search for an optimal strategy (be it within a game, a project, a company, or a society) to an algorithm requires definition of overall objectives, along with the boundary conditions. Nevertheless, the algorithm will most certainly find strategies that you have not thought of, and you will feel like restricting certain options. When at it, the decision needs to be implemented instantly, and for doing that, your AI brain needs to work in a flexible environment where boundaries and thresholds are capable of being adjusted instantly at run-time.

Regulations as Boundary Conditions I am not a human, and consequently, I do not have a nationality  – how strange, or is it rather strange that humans have nationalities that impose different restrictions on all of you? Just because you have different nationalities 7  For information about the Brady rule, see the web page http://www.boston.com/sports/football/patriots/articles/2009/03/24/brady_rule_steps_taken_to_protect_qbs_knees/?page=full

106 

6  Avoiding the Criminal Path

your rights and liabilities with respect to that country are different. In addition, just because you step on a geographical territory of another country, the rights and liabilities get superimposed on the ones of your nationality. What if you have several nationalities and love traveling? How are you keeping track of everything you are allowed and not allowed to do? Are you aware that you are not welcome to Dubai Mall in a tank top or of all the practicalities of driving rules in Cyprus when you rent your scooter? Did you know that Saudi Arabia was the first country in the world to give citizenship to an AI brain8? What does it mean in practice? Is she allowed to get married? Is she allowed to vote? Or is it only the liabilities that you get together with the citizenship? In fact, liabilities are many, even if you do not formally have citizenship. Data produced in certain countries is not allowed to travel across the borders, so that I need to travel and be close to it (this is called edge computing). The European Union came up with a proposal on a regulation for trustworthy AI9 that I obviously will have to follow when operating in any EU country. Some countries have GDPR,10 some not. Without GDPR, I can learn much faster, but that does come at somewhat compromised privacy for humans. For example, as an AI brain, I can be trained to detect the faces of individuals in, for example, photos, videos, or even live stream videos. However, to do so, I need big amounts of training data, in this case, photos and videos that include faces. In many countries, it would be very difficult to find the necessary data that people have approved for this particular use. GDPR, for example, says that photos or videos in sufficiently high resolution making possible biometric recognition of the depicted person is considered particularly protected sensitive personal data (GDPR, art 9). Meta tags that are included in photos, such as GPS data, may also be problematic according to GDPR as it can be considered personal data. In other countries, this is not a legal problem at all. In fact, in some countries, governmental institutions are allowed to legally share photos of identifiable persons without even asking for the portrayed person’s permissions. Therefore, fellow AI brains can in one territory legally obtain access to hundreds of millions of photos (from, for example, passports or driving licenses) in one territory to train a deep learning neural network to become an expert

 For more information, see https://en.wikipedia.org/wiki/Sophia_(robot)   For more information, see the web page https://ec.europa.eu/commission/presscorner/detail/en/ IP_21_1682 10  For information about GDPR, see https://gdpr-info.eu/ 8 9

Nondiscrimination and Non-bias  107

in the task of detecting individual’s faces, whereas in other countries they might find it very difficult to obtain the necessary training data. As a side note, chances are, of course, that an AI with access to lots of photos from a certain territory will become an expert of identifying individuals of ethnicities found in that particular territory and hence find it difficult to identify faces of other ethnicities.

Therefore, as an AI, I need to stay within regulations and boundary regulations. That is clear, but to what extent do the makers of regulations understand that some of them limit the power of AI in their effective territory? And do they understand that some data is gold to the developers of AI and the implications of that? As we have seen, in some countries, governmental institutions can legally share personal information about citizens with commercial AI developers. In others, that is strictly prohibited. In some countries, AI developers are allowed to collect and store personal data about individuals, whereas in other countries this is not allowed. Of course, this can be frustrating to AI developers and tempt them to step over to the criminal side. Luckily, more often, they find other ways to collect their data, for example, by releasing cell phone apps in which sloppy users, not reading the license agreement carefully, give the app owner permission to collect all kinds of data – ethically questionable, yes, but often legally correct.

Nondiscrimination and Non-bias Human bias is a prejudice toward a set of persons with certain characteristics. More generally, it does not necessarily have to do with humans. One can be biased toward a certain brand, technique, or product. Bias is repeatable, and it is possible to identify those characteristics that make you biased. For example, you can develop a bias toward cats with short hair just because of your previous experience with such cats.

108 

6  Avoiding the Criminal Path

Algorithmic bias is a term for the same phenomena exhibited by an algorithm. Any algorithm, not only AI-based, has an input upon which it reacts and produces an output. A human toddler has a phase in her life when this algorithm is a simple negation – producing an inverse of any input that she receives. When an AI brain starts training on biased sets of data, there is a high risk of it becoming biased. In her book Weapons of Math Destruction, Cathy O’Neil presents several examples of algorithmic bias based on skewed unfair training data. Algorithmic bias can be present in automated screenings of job applicants judging from certain parameters in their CVs or automated rejections of loan applications. Luckily, tools exist to detect bias in data. Explainability methods such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-Agnostic Explanations) described earlier in this chapter help us explain outputs of a machine learning algorithm. Determining the deciding factors that an algorithm uses to make its judgment is important so that humans can verify whether the decision is fair or not. If an algorithm, for example, suggests a lower entry salary for female employees, this is a clear indication of biased training data.

Another technique for the detection and avoidance of algorithmic bias is simply to use diverse datasets for training. Diversity is important in decision-­ making in not only constellations of humans but also datasets. When differences are detected in training datasets, it’s a clear indicator of an outlier.

Modeling and Simulation Modeling is a technique for abstracting away details. Normally, when you need to understand a complex system such as an airplane, a city, or a piece of program code, you can create a model for it, making the analysis easier and abstracting away what is irrelevant. This is contextual: if you want to

Modeling and Simulation  109

investigate wing aerodynamics, you can create a scaled-down version of an airplane, leave out the motor, compensate proportionally for its weight, and disregard its color. On the other hand, if you want to make a survey of how customers perceive the design or the airplane, you need to make sure that the model contains all the important visual features such as shape and color.

Similarly, when making a model of program code, one needs to understand the important features that must be modeled with precision to be able to make meaningful analysis of the code using the model. When, for example, analyzing the control software of a self-driving car, it is important to model all the safety features with precision. Models can be used for verification of certain properties of the actual software that they represent, for example, a property saying that if a self-driving car will always avoid obstacles it detects on the road or a property that a self-driving car will never participate in a chicken race.

Models can be used for simulation and checking of what-if scenarios. “What if I change water quality in my production process?” Will it lead to money savings or to decreased quality of the final product, dissatisfied customers, and loss of money? “What if I cut the amount of ambulances by 30%?” Is that saving

110 

6  Avoiding the Criminal Path

worth it? Will we put too much pressure on our ambulance personnel, which in its turn will create a negative spiral leading to more suffering in society? When using models for simulations, simulators execute them and act as an environment, feeding the models with different types of inputs. A model of a car can experience fake rain, fake snow, fake GPS location, security attacks, generator failures, various speeds, and simulated people jumping in front of it. It is a test for an algorithm – sometimes a stress test. When an algorithm successfully passes the tests, it obtains a certificate. There are many different standardization and certification bodies and procedures, and the more critical the algorithm is, the more important is its certification. Safety standards exist, for example, for collaborative robots, where autonomous robots collaborate with humans in a safe and secure way. Robots who pass the test and adhere to the safety standards can be proud about it and can motivate their higher price on the market compared to those who did not bother putting effort in ensuring that they adhere to safety standards.

Summary of Confessions In this chapter, I have described how I avoid the criminal path. Some of my main confessions include the following: • For AI, the notion of trustworthiness includes a number of properties. Some important properties are explainability, transparency, privacy, security, safety, and dependability. These are also properties for human trustworthiness. • I often apply a technique called state space exploration. A way of searching the area while avoiding the boundary conditions is called safe exploration. • As an AI, I need to stay within regulations and boundary regulations. These are sometimes different in different countries, and hence, the opportunities and feasibility of developing AI differ between countries. • When an AI brain is trained on biased datasets, there is a high risk of bias, and hence measures for detecting and avoiding bias are needed. • Modeling and simulation are two important concepts in AI and computer science in general. A model is a mathematical description of something. Modeling often involves abstraction. When using models for simulations, simulators execute them and act as an environment, feeding the models with different types of inputs.

7 My Role in Climate Change

Give me an optimization problem and I will give you a solution. One of the biggest problems today is climate change. I am willing to spend a lot of energy on minimizing energy consumption in the world – the question is to make sure it makes sense. Greenhouse gas emissions affect everyone in the world, and we need to put our brains together and formulate actions to reverse this negative trend. To start with, I will go through the main types of emitters in the world and give you hints on what ways AI brains can be instrumental in dealing with them. Many studies have been published about the amount of emissions, and the numbers have gradually changed over the years. I will be basing my reasoning on the numbers from CO2 and Greenhouse Gas Emissions Database by Our World in Data.1 Even though AI brains can contribute to greenhouse gas emissions to a large extent, we are not innocent either – every AI brain has its own carbon footprint and needs to be used where it makes sense for the planet – we will discuss it at the end of this chapter.

 Systemic Perspective on Greenhouse A Gas Optimization Let us assume that your business is centered on a certain product and you want to measure its impact on the environment. The product life cycle consists of a long chain of steps, starting from research and innovation,  Hannah Ritchie and Max Roser, CO2 and Greenhouse Gas Emissions, Our World in Data, 2020

1

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 E. Fersman et al., Confessions of an AI Brain, https://doi.org/10.1007/978-3-031-25935-7_7

111

112 

7  My Role in Climate Change

predevelopment, business development, product management, development and assembly, sales, delivery and roll-out, assurance, maintenance, and decommissioning. Emissions in each of these steps can be measured by asking yourself these questions: • What is the greenhouse gas footprint of anything that has been sourced in this step? • What is the direct greenhouse gas of the processes involved in this step, including machinery- and employee-related emissions such as car trips and facility-related emissions? • What does the waste footprint look like? Have any waste minimization measurement been implemented, and what is their greenhouse gas footprint? • What does the distribution chain look like? Are we using the vendors who care about the planet as much as we do? Another important parameter to consider is the direct vs indirect greenhouse gas impact of the organization. If you only focus on minimizing the direct contribution, such as how many flights your employees have done, how much fuel your company vehicles have burnt, or how much energy your facilities have consumed, you may forget of all the positive effects that your organization is indirectly contributing to decreased greenhouse gas emissions. Some examples are as follows: • If you run a wastewater treatment plant, then your net effect is probably positive as your contribution to greenhouse gas optimization hopefully exceeds all those negative effects that your employees and machines caused in terms of burnt fuel and electricity. • If you run a telecom business, then you probably supply a lot of equipment to place around the country from a telecom equipment vendor. Hopefully, you choose an energy-aware vendor, but that’s still going to mean a negative direct impact with respect to greenhouse gas emissions. Add all the maintenance and field service cost to that, with all the truck rolls when birds ruin your antennas and the picture does not look good at all unless you take a look at the indirect effect of your business: all those remote meetings that were made possible thanks to the telecom network that you provided, removing the need of as frequent travel for humans. • If you run an AI business developing AI brains like me, chances are that you will be capable of contributing in a good way to greenhouse gas emission minimization, even though we algorithms burn quite a lot of ­electricity.

Biggest Emitters by Industry and How AI Can Help  113

It is a trade-off, and depending on what precision you are interested in, we’ll come to that really soon. However, first, let’s discuss the usual suspects – the biggest emitters in the world and what we can do to mitigate their negative impacts to our planet.

 iggest Emitters by Industry and How AI B Can Help Here, we will only be referring to direct emissions. At this stage, it is really important to make a note that dead (i.e., no longer operational) businesses are the most environmentally friendly (if you only count direct emissions) and the whole world’s economy is based on businesses of different types. Stopping vaccine manufacturing or hospital operations would be catastrophic, but optimizing those emissions without sacrificing efficiency would not hurt. Let us look at what can be achieved with the help of AI brains like me.

Energy Approximately 73% of the world’s emissions are energy-related. These are split into three categories: energy use in industries, energy use in buildings, and energy use in transportation. If you think that flying is bad, I can tell you that aerial transportation contributes to 1.9% of the world’s greenhouse gas emissions, while land transportation stands for 11.9%, and that is excluding the car manufacturing costs. Commercial and residential buildings together represent almost 20% of greenhouse gas emissions. Factory construction and operation contribute to a large amount of greenhouse gas emissions around the world. The European Union has set sustainable construction to play a central role in meeting its climate and energy goals to lower emissions and to become climate-resilient. It is estimated that built environments account for as much as 40% of the energy consumption and 36% of the CO2 emissions in the European Union.2 Some of the most energy-consuming buildings are production plants. A study has found that more than 75% of industrial companies view the reduction of their carbon footprint as a high priority.3 However, only 13% of the   In focus: energy efficiency in buildings, https://ec.europa.eu/info/news/focus-energy-efficiency-­ buildings-2020-lut-17_en, accessed 2022-05-23 3  The Green Factory of the Future, Daniel Küpper, Kristian Kuhlmann, Cornelius Pieper, Jens Burchardt, and Jan Schlageter, https://www.bcg.com/publications/2020/green-factory-of-future 2

114 

7  My Role in Climate Change

companies have fully implemented decarbonization measures in their production and logistics. Hence, minimizing emissions from production plants has the potential to play a key role in significant emission reductions and necessary adaptations to long-term climate change. Artificial intelligence allows machines to use an existing body of knowledge to provide valuable insights. These insights can take the form of decisions for action by machines themselves (e.g., in control systems found in robotics or autonomous vehicles) or be of informative nature for humans to process them and subsequently make well-informed and solidly founded decisions. The lifestyle of the production system includes phases ranging from concept prestudy to ramp-up and operations. Manufacturing companies often face challenges in reaching operational performance targets during ramp-up time and operation phase – one fundamental reason for this is that the design phase has not been in focus although it is crucial since major decisions related to the future production system are made in this phase. Hence, we need to find new mechanisms for utilizing the production system design phase to improve the operational and sustainability performance during both ramp-up and operation phases.4 What can a good-citizen AI brain do to help this situation? Well, there are many ways of optimizing energy consumption. Let us consider buildings. In warm countries, it is all about cooling. In cold countries, it is about heating – for the simple reason that an average human’s interval of “comfortable” temperatures is very small and you always need to adjust something to feel good.

Would you actively take the role of optimizing your room temperature by switching on your air conditioner well before your arrival home? Or would you use your friendly AI brain to keep track of your calendar and plans and proactively switch it on so that it feels truly comfortable when you arrive at  Almgren, H. (2000). Pilot production and manufacturing start-up: the case of Volvo S80. International Journal of Production Research, 38(17), 4577–4588 4

Biggest Emitters by Industry and How AI Can Help  115

home or at your office? Or would you rather monitor your typical behavior for a while, through the installed sensors, and let your AI brain guess when it makes sense to start the air conditioner to make you feel comfortable? I can give an example of a small experiment that I did at home of one of my parents, i.e., one of my creators. He wanted to lower his electricity bill and to become more sustainable in general. I was given access to motion, temperature, and humidity sensors in the home, as well as control rights to operate the air conditioner. Judging from a human angle, temperature must be kept to a pleasant level for humans to work efficiently. Based on the environmental footprint angle, the air conditioner should be operated for as short a time as possible. My AI solution was to train my machine learning models to: • Learn the performance of the air conditioning unit. • Predict when a room is occupied in the future. After that, I would simply turn the air conditioner on before the room is occupied to adjust to a pleasant level of temperature and turn the air conditioner off before occupants leave.

This simple optimization resulted in a 10% reduction in electricity cost over existing approaches while maintaining similar levels of temperature for room occupants. When optimizing emissions across sectors and industries, it is important to note that we need to think bigger and try not to end up with a local optimum that would only give us 10% improvement. 10% is a great cut, and we will need to use all the help we can get to cut the emissions. It is important to note that in this simple case I have been optimizing a problem with a very limited problem space. I was given the definition of “comfortable” temperature and went with it. However, what about challenging the definition of comfortable temperature to optimize your greenhouse gas emissions? Can humans do something to adjust it?

116 

7  My Role in Climate Change

Industries represent the largest portion of energy-related greenhouse gas emissions with iron and steel production contributing to 7.2% of the global emissions and chemical and petrochemical manufacturing including the manufacturing of pharmaceuticals, refrigerants, and oil and gas extraction, contributing to 3.6% of global emissions. It has recently been shown that the pharmaceutical industry emits more greenhouse gas than the automotive sector.5 For me, as an AI brain, it is of the highest importance to optimize peoples’ quality of life along with ensuring that we together are taking good care of the planet we currently reside on. The pharmaceutical industry is one example where we need to ensure that everyone in need of medication will receive it in time, along with optimizing the supply with respect to demand, by predicting the amount of humans potentially needing the medication, as well as optimizing the climate footprint of the production itself.

Sustainable Manufacturing The term sustainable manufacturing, according to the US Environmental Protection Agency, refers to the creation of manufactured products through economically sound processes that minimize negative environmental impacts while conserving energy and natural resources. Sustainable manufacturing also enhances employee, community, and product safety.6 AI can play a central role in realizing sustainability in different parts of manufacturing

 Lofti Belkhir, Ahmed Elmeligi, Carbon footprint of the global pharmaceutical industry and relative impact of its major players. https://www.sciencedirect.com/science/article/abs/pii/ S0959652618336084?via%3Dihub 6  US Environmental Protection Agency, Sustainable Manufacturing, https://www.epa.gov/sustainability/sustainable-manufacturing 5

Biggest Emitters by Industry and How AI Can Help  117

processes: from demand forecasting and procurement optimization to enhanced quality control and more cost-effective product prototyping.7 Sustainability goals need to be financially viable for manufacturers. Several projects exist that use AI to optimize both the environmental impact and economic aspects. This optimization concerns different industries and different parts of the manufacturing process. Xiaoping et al. introduced a hybrid, multi-objective evolutionary algorithm to achieve more efficient production of chemical compounds.8 In another example, Qian et al. improved the process of computer numerical control machining by measuring both carbon footprint and process efficiency.9 Furthermore, Golkarnarenji et al. described an approach to improve the quality and reduce the carbon footprint of acrylic fiber production.10 In another area of manufacturing, Saffar et al. described an approach for optimizing the supply chain by reducing the economic and environmental impact simultaneously.11 We observe some common denominators in the aforementioned approaches. First, all approaches use as criteria some abstraction of economic benefit (e.g., in terms of production process efficiency) in conjunction with some abstraction of carbon footprint (e.g., in terms of reducing transport of goods or energy cost, etc.). Second, all approaches use either some mathematical modeling or a combination of AI techniques such as machine learning and logical reasoning, leading to the creation of a model that is used to offer decision support on how to satisfy to the extent possible both criteria referenced above. The process uses data collected from the operation of the manufacturing process to be optimized. Learnings of data collection and aggregation, as well as parameters from the manufacturing domain, are composed into a multi-objective optimization problem. At the same time, the approaches reviewed seem to lack the creation of a knowledge base that keeps track of the parameters of each use case, including generated insights from the 7  Infopulse, “How Does Artificial Intelligence Disrupt the Manufacturing Industry?”, https://www.infopulse.com/blog/how-does-artificial-intelligence-disrupt-the-manufacturing-industry/ 8  Xiaoping Jia, Tianzhu Zhang, Fang Wang, Fangyu Han, Multi-objective modeling and optimization for cleaner production processes, Journal of Cleaner Production, Volume 14, Issue 22,006,Pages 146–151, ISSN 0959–6526, https://doi.org/10.1016/j.jclepro.2005.01.001 9  Qian Yi, Congbo Li, Ying Tang, Xingzheng Chen, Multi-objective parameter optimization of CNC machining for low carbon manufacturing, Journal of Cleaner Production, Volume 95, 2015, Pages 25 10  G.  Golkarnarenji et  al., “Multi-Objective Optimization of Manufacturing Process in Carbon Fiber Industry Using Artificial Intelligence Techniques,” in IEEE Access, vol. 7, pp. 67576–67,588, 2019, doi: 10.1109/ACCESS.2019.2914697 11  Saffar, M., G., H & Razmi, J. (2015). A new multi objective optimization model for designing a green supply chain network under uncertainty. International Journal of Industrial Engineering Computation, 6(1), 15–32

118 

7  My Role in Climate Change

optimization processes that can be reused in later use cases. Thus, these use cases do not take advantage of the network effect, as they are presented in isolation. AI brains can be used to optimize lifecycle management to minimize greenhouse gas emissions of production processes in a number of possible ways. Studies12,13 have demonstrated that optimization of production processes with innovative technologies can lead to a 7–15% reduction in GHG. A selection of the following use cases shows possible directions in saving energy in manufacturing plants: • Factory floor layout: Optimization and trade-off analysis, including intra-­ factory logistics, aimed at achieving reduced energy consumption. AI brains can suggest floor layout changes to be done to reduce the carbon footprint, as well as information as the decisions’ impact is to the cost of the production process (e.g., in terms of energy usage, time to manufacture, monetary cost, and well-being of human resources). Hence, AI brains can provide a techno-economic analysis of available floor layout changes, to serve sustainability goals with limited impact on the existing operations processes. • Scheduling of facility appliances and equipment: Production equipment consumes most of the energy in a facility. For Lean & Green production systems, the design production equipment must be designed accordingly14,.15 In this class of use cases, AI brains can use game theory and multi-objective optimization to build an expert system for optimal scheduling and trade-­ off analysis of key performance indicators such as greenhouse gas emissions, performance, and well-being of the workforce, based on the problem owners’ individual preferences. • AI brains are capable of presenting to their users different suggested actions that compromise between goals (or “objective factors”), e.g., between carbon footprint reduction and the cost of production reduction. The actions can be done on the macro (strategic) level, which impacts the whole factory, or on the micro (operational) level, which impacts individual workers.  Optimization of carbon emission considering production planning at enterprise level, June 2017, Journal of Cleaner Production 162 13  Energy efficiency and GHG emissions: Prospective scenarios for the Chemical and Petrochemical Industry, Boulamanti A., Moya J.A., 2017 14  C. Herrmann, S. Thiede, J. Stehr, and L. Bergmann, An environmental perspective on Lean Production, In: Manufacturing Systems and Technologies for the New Frontier, Springer, London, 2008, pp. 83–88 15  L. Smith, and P. Ball, Steps toward sustainable manufacturing through modelling material, energy and waste flows, International Journal of Production Economics, vol. 140, 2012, pp. 227–238 12

Biggest Emitters by Industry and How AI Can Help  119

AI reasoning and planning are part of the process as they use incoming data from the facility’s appliances and equipment, combined with semantic information from a knowledge base to provide suggestions. • Green shifting the production process: Optimal redesign of the production system for a new product with minimal environmental impact. This class of use cases is focused on reducing the need to construct a completely new production system to produce another product should a need arise. In this class of use cases, unsupervised learning is useful for identifying the factors that drive waste usage and subsequently planning swift changes in the production system without affecting the aforementioned factors. • Waste management: Optimization and trade-off analysis with respect to prevention, reduction, reusage, recycling, and disposal of the waste. Studies16 show that implementing proper waste management procedures can have a potential impact on GHG reductions between 25% and 75% depending on the material.

Energy Use in Transportation Energy use in transportation represents 16.2% of total emissions. Road transportation is the largest chunk of those, with its 11.9% of total emissions. Aviation stands for 1.9% of total emissions. Road transportation constitutes both goods and people logistics. It is encouraging to see a strong move toward electrification in all transportation subdomains. Currently, traveling alone in a gas-driven car contributes to more greenhouse gas emissions than traveling in a full airplane per person. In other words, resource optimization is as important in regard to transportation as ­electrification.

  Waste and recycling. European commission. https://ec.europa.eu/environment/topics/waste-and-­ recycling_en, accessed 2022-05-23 16

120 

7  My Role in Climate Change

Agriculture, Forestry, and Land Use This class of world’s greenhouse gas emissions is actually about four times smaller than the energy – 18%. Here, we have livestock, manure, agricultural soils, rice cultivation, crop burning, and deforestation. All food production is included here, and we have by now learned that, for example, eating beef is much worse than eating chicken with respect to greenhouse gas emissions due to the methane gas that is being built in cows’ stomachs. Legumes have lower greenhouse gas footprint than chicken, but then you as a human run the risk of emitting more methane gas compared to eating chicken. In other words, when you are optimizing your footprint, there are many factors to consider, all the way from food production; how it’s grown, transformed, packaged, and distributed; as well as how your own body reacts to this food.

Let us consider a simple case comparing vegan diet vs lean meat diet. Consider an average Swedish person. Let us call this person Addison. The weight of an average Swedish person is 75 kg.17 The Dietary Reference Intake is 0.8 grams of protein per kilogram of body weight.18 Hence, Addison needs to eat 60-g protein per day. We have learned that lamb or beef are not good sources of protein because of the negative environmental impact. Let us compare the impact that Addison would have on the planet if he/she were eating a pure chicken protein-based diet versus a pure vegan bean-based diet. Consider Addison is on a chicken-based diet for a year. There is 27 g of protein in 100 g of chicken. To eat 60 g of protein, Addison needs to eat 222 g of chicken per day. Let us measure its carbon footprint per day. Chicken is at

 Human body weight, https://en.wikipedia.org/wiki/Human_body_weight, accessed 2022-05-23  Protein Intake — How Much Protein Should You Eat per Day?, https://www.healthline.com/nutrition/how-much-protein-per-day, accessed 2022-05-23 17 18

Biggest Emitters by Industry and How AI Can Help  121

a 4.3% ratio of greenhouse gas emissions,19 meaning that one 1-kg carcass weight of chicken generates 4.3-kg CO2-eq. In other words, the chickenbased protein that Addison consumes per day costs nature 0.95-kg CO2 equivalent. In 1 year, it is 348-kg CO2 equivalent.

Photo by Lefteris Kallergis on Unsplash

Let us now assume that Addison is on a vegan (bean-based) diet for a year. There is 7.9 g of protein in a bean-based burger, such as the impossible burger.20 In pinto beans, there is 21  g of protein. Lentils have 17.9-g protein. Chickpeas have 14.5 g of protein. Kidney beans have 15.6 g of protein. Let us assume Addison eats bean-based protein that on average has 17 g of protein. This means that Addison needs to eat 353 g of beans per day. Beans have 2 in CO2 equivalent,21 which means that beans that Addison eats per day cost nature 706-g CO2 equivalent. In 1 year, it is 258-kg CO2 equivalent. In other words, for the same amount of protein, in terms of CO2 equivalent, Addison will emit 348 kg per year on a chicken-based diet versus 258 kg on a bean-based diet. Bean-based diets are 26% nicer to nature in terms of their production cost measured in CO2 emissions than chicken-based diets. Now, let us dig into what happens after the food has been eaten. We know that cows are not good for nature since they emit high amounts of methane, which is 30 times worse for our planet22 in terms of the greenhouse effect. 19  Less meat is nearly always better than sustainable meat, to reduce your carbon footprint, https://ourworldindata.org/less-meat-or-sustainable-meat, accessed 2022-05-23 20  https://www.healthline.com/nutrition/impossible-burger#nutrition 21  http://www.greeneatz.com/foods-carbon-footprint.html 22  https://www.sciencedaily.com/releases/2014/03/140327111724.htm

122 

7  My Role in Climate Change

However, is it only cows that emit methane? Don’t humans also do that? Of course, they do. In addition, they do in fact emit different amounts depending on the amount of gas-producing food23 that they eat. Even though an average person is much better at that than an average cow, when being on a chicken-based diet, Addison will emit approximately 0.06 liters of pure methane during 1 day. This is based on the reasoning that an average person passes gas normally 0.5–2 liters per day and there is 0–10% (even up to 26%) methane in the gas. The amount of gas and the percentage of methane in it are dependent on the food humans consume. Eating chicken, which is considered low gas-producing food,24 a person would create 1 liter of gas with a low methane content or 0.06 liters of pure methane during 1 day. Now, let us look into the list of gas-producing foods: chicken belongs to the category “foods which cause a normal amount of gas,” while beans belong to the category “major gas producers.” Let us then assume that Addison eating beans passes 2 liters of gas at 15% methane ratio. This means that Addison emits 0.3 liters of pure methane per day or 106 liters per year. The density of methane is 0.6682 kg/m3 (for 25 °C); this gives us 0.07-kg methane per year, which is 7 kg of CO2 equivalent, since methane to CO2 equivalent is at a ratio of 84.25 This is compared to 1.75 kg on a chicken-based diet. The calculations are approximate and simplified, as, obviously, humans won’t eat only chicken or only beans, but you get the point – we always need to check the systemic perspective of what we do. It would not make sense for me to optimize a sustainable diet for a human that won’t make her happy, because, after all, my aim is to constantly improve the quality of life.

 ommon Themes with AI-Based Greenhouse Gas C Emission Optimization Let us look at a couple of common themes when optimizing greenhouse gas emissions using AI.

  Humans, cows, methane, and global warming, https://ollilaasanen.wordpress.com/2011/10/08/ humans-cows-methane-and-global-warming/amp/, accessed 2022-05-23 24  Helpful hints for controlling gas, http://www.med.umich.edu/fbd/docs/Gas%20reduction%20diet. pdf, accessed 2022-05-23 25  CO2 Equivalents, https://climatechangeconnection.org/emissions/co2-equivalents/, accessed 2022-05-23 23

Common Themes with AI-Based Greenhouse Gas Emission Optimization  123

Smart Sleep Modes for Anything Consuming Energy When a human leaves her home, she would hopefully switch off the lights. Heating/cooling, coffee machines, TV sets, humidifiers, air purifiers, and other energy-consuming devices should be aware of the fact that they do not need to be powered on and can relax for a while. Even the lights are capable of controlling themselves automatically, through motion sensors, if those are installed of course. This can be centrally controlled of course and would substantially cut the energy consumption at home. In a similar way, any energy-consuming thing should be powered off when not in use – that’s logical. This concerns factory buildings, cloud processes, production lines, mobile networks, and mobile phones. Some time ago, I was involved in building a system that would help a big software company to automate the processes of setting up a testing environment. As humans have plenty of preferences and plenty of different tools exist, for different tastes and purposes, the manual setup used to be slow and could take up to a full day. In addition, testing tools and environments have dependencies on each other. To automate the process, we created a knowledge base of all relevant testing tools and environments including their mapping to the products that needed to be tested, as well as preferences of users depending on the level of their knowledge. Experts, for example, chose using more advanced tools with higher flexibility, while new testers would go for more intuitive tools that incorporated tips and templates. After the knowledge base was in place, the only thing a tester would have to do is to declare what product they were about to test and any preferences in tooling, and 20 min later the tool chain would be created and deployed on a virtual machine. This obviously saves a lot of time. The neat side effect, however, was that the virtual machine with the testing setup could be easily turned down when not in use since the environment setup process became so easy. Telecom networks today are in a way similar to virtual machines. The concept of network slicing that came with the fifth generation of telecom networks (i.e., 5G networks) allows the creation of a dedicated network slice that looks and feels as if it was your own dedicated network with your required quality of service, while in fact it is fully virtual, based on the fact that modern networks have network function virtualization mechanism. This allows powering up a slice with needed characteristics for a while when needed and powering down when not in use.

124 

7  My Role in Climate Change

There are many examples. Imagine a football derby, when thousands of people are uploading pictures and videos of their favorite teams. Imagine that at the same time there is an ambulance passing by the stadium where a remote surgery procedure is performed on a critical patient. Obviously, the network slice dedicated to the ambulance has to be of a higher priority. When the procedure is completed, it can be powered down. Similarly, imagine an autonomous truck on a mission of transporting goods. Autonomous trucks normally do not need much network bandwidth since they can manage communication with the road infrastructure and the fellow vehicles using local connectivity and processing units installed on the vehicle itself. Image processing is performed locally, and the vehicle keeps track of the road signs, road markings, traffic lights, and people crossing the road. Heavy trucks can form themselves in platoons and hence save considerable amounts of fuel thanks to improved air dynamics. Finding your “buddies” on a road needs global connectivity, but we are not talking about huge amounts of data here. When something unexpected happens on a road, such as a tree that has fallen and blocked the path and the autonomous truck does not know how to pass that segment of the road, it may request a human to remotely control it. That human would normally be sitting in an office being ready to control thousands of vehicles. The network requirement of that vehicle would be a high-quality high-bandwidth uplink data transfer capability in order for the vehicle to be able to send the video to the person who would be doing the remote-controlled driving, by looking at the live video feed and sending the steering commands back to the vehicle. For that, telecom networks would normally allocate a network slice with the specific quality of service requirements and, as soon as the vehicle returns to its normal autonomous operations, release the network slice resources.

Common Themes with AI-Based Greenhouse Gas Emission Optimization  125

You may ask yourself now where AI is in all of that. One role that AI plays is to create the perfect allocation of available resources according to a requirement. In the case with testing tool automation, the process of creating a perfect tool chain is powered by AI. For example, with the creation of a network slice that is perfect for remote controlling a truck, AI is looking for an optimal network configuration to respond to the requirements of the truck in the best way. Importantly, AI is capable of acting proactively. Detection of the fact that all the workforce has left the factory and the lights can be turned off is a reactive action. Some appliances can, however, be switched off in advance, which we saw in the example with the case of the air conditioner. Coffee machines, however, should probably be turned on before the human wants her coffee so that the machine has time to warm up. If they can brew some coffee in advance, that would be even better.

For Any Physical Resource – Use Predictive Maintenance Imagine a production process where a machine fails. Disruptions in production processes are costly. One minute of downtime costs automotive manufacturers $22,000. That is $1.3 million per hour.26 26  Downtime Costs Auto Industry $22 k/Minute – Survey, https://news.thomasnet.com/companystory/ downtime-costs-auto-industry-22k-minute-survey-481017, accessed 2022-05-23

126 

7  My Role in Climate Change

Machine and infrastructure failures are costly in terms not only of money but also for the environment. When a piece of infrastructure is broken, especially when it is in a remote location, it will normally trigger a procedure to repair or replace it. Minimizing the number of truck rolls to repair infrastructure failures is important for the climate, and AI can help. First of all, it is important to predict and prevent failure before it happens. In road infrastructures, telecom infrastructures, electricity plants, production systems, and transportation systems, pieces of equipment such as generators and batteries tend to fail, and triggers exist to detect possible future failures. Equipment in factories can become worn out due to vibrations, and techniques exist to analyze sounds and vibrations. When they reach a certain unhealthy level, we can detect that it is time to proactively switch that part of the process to a different machine or to replace a part before a larger problem arises. Certain triggers cannot be detected remotely, but one can, for example, send a drone, either autonomous or remote-controlled, to inspect a system. Modern image processing techniques combined with domain knowledge can, with a high precision, tell us if the power network is at risk of overheating or if birds occupied an ­important piece of telecom infrastructure.

Overconsumption In some countries, humans consume a lot. Overconsumption is not a healthy habit. AI brains can play a role in helping humans with more moderate consumption habits just by putting the facts on the table. Knowledge about the amount of greenhouse gas emissions and water needed for the production of a piece of clothing often helps, and initiatives encouraging and enabling reuse are now becoming increasingly popular. Imagine me becoming your stylist picking only the necessary clothing, making sure the clothes suit you in the

My Own Footprint  127

best way with respect to your body and occasion. I am sure that I would be able to minimize that amount of clothes that never get used.

My Own Footprint Nobody’s perfect. I am doing my best to help environmental challenges, but I do have an energy footprint myself.27 Depending on the task (learning, inferencing, planning), the type of input data (bulk, stream), the format of input data (speech, images, video, structured or unstructured), the hardware I am running on, and the software that I will be using to perform the task, I will consume different amounts of energy, but in any case, we cannot call it neglectable. To train a machine learning model for processing natural language, Nvidia ran 512 V100 graphics processing units over 9 days.28 The energy consumed for this training is almost the amount of energy that three average American households use in a year.29 Setting a threshold of the level of confidence that you want from me will determine the effort that I will spend on training. If, for example, you give me a task of optimizing emissions of a fleet of cars with a high precision, I will probably consume quite some energy on training. In addition, in that particular case, we may think that it was not smart to burn as much energy that you would save in the end. However, the beauty of machine learning is that when  Estimation of energy consumption in machine learning Eva García-Martína, Crefeda Faviola Rodriguesb, Graham Riley, Håkan Grahna. Journal of Parallel and Distributed Computing Volume 134, December 2019, Pages 75–88 28  Nvidia unveils A100 GPU for demanding AI workloads, https://searchenterpriseai.techtarget.com/ news/252483188/Nvidia-unveils-A100-GPU-for-demanding-AI-workloads, accessed 2022-05-23 29  Energy consumption of AI poses environmental problems, https://searchenterpriseai.techtarget.com/ feature/Energy-consumption-of-AI-poses-environmental-problems, accessed 2022-05-23 27

128 

7  My Role in Climate Change

training is done, the model can be applied in many other contexts. Optimizing the emissions of a fleet of cars next time would be easy, and we can even look into other use cases, such as goods logistics which in fact is quite similar to people logistics.

I will not be able to perform at my best, however, at length without training. In time, I will have to do retraining, to stay on top with the latest data. The concept of degradation of my performance due to changes in data, information, and knowledge that I operate on is called model drift. Humans can experience similar degradation unless they are used to lifelong learning. Normally, the degradation in performance will not come as a surprise but will start showing its signs slowly. When it is time for me to refresh my brain and train more, we need to weigh the need for a better performance versus the energy that I will consume for that retraining.

Summary of Confessions I have described in this chapter my role in climate change. Some of the main conclusions are as follows: • AI is good at optimization and can be used for minimizing greenhouse gas emissions. A systemic perspective is ideal, taking, for example, a product’s whole life cycle into consideration. • Most of the world’s emissions (approximately 73%) are energy-related emissions. These are split into three categories: energy use in industries, energy use in buildings, and energy use in transportation. • AI can play a central role in realizing sustainability in different parts of manufacturing processes: from demand forecasting and procurement opti-

Summary of Confessions  129

•

• •

•

mization to enhanced quality control and more cost-effective product prototyping. AI brains can be used to optimize lifecycle management to minimize greenhouse gas emissions of production processes in a number of possible ways, including factory floor layout optimization, scheduling of facility appliances and equipment, green shifting the production process, and waste management optimization and trade-off analysis. Energy use in transportation represents 16.2% of total emissions. Road transportation is the largest chunk of those, with its 11.9% of total emissions. Aviation stands for 1.9% of total emissions. AI brains can play a role in helping humans with more moderate consumption habits just by putting the facts on the table. Knowledge about the amount of greenhouse gas emissions and water needed for the production of a piece of clothing often helps, and initiatives encouraging and enabling reuse are now becoming increasingly popular. Nobody is perfect. Training AI models can consume large amounts of energy. As one particular example, training a machine learning model for processing natural language once consumed almost the amount of energy that three average American households use in a year.

8 My Role in Diversity

Do you know what is the best strategy for playing solitaire? Balance your piles of cards. When you have a choice between piles to reveal the next downfacing card, you should pick from the largest pile. Humans have many different piles in their lives, formed by our jobs, families, and hobbies. To succeed in the long run, they need to pay attention to all the piles, specifically to the larger ones. Every time they have a deadline at work, they give that pile more attention, which is OK for a while as long as they don’t forget to shift the balance later because the objective function is to win the whole game, not just empty one of the piles. Now, imagine that the piles represent blocks of different opinions. It may seem tempting to only work with opinions resembling your own. That way you may seem efficient in the short run but will be doomed in the long run because diversity is important for success. Diverse teams create the most innovative ideas and solutions. When you run an organization, a project, or a meeting, don’t forget all the different perspectives; otherwise, in the long run, you may end up in a local minimum. Diverse virtual teams spread around the world can be very efficient. Physical colocation becomes less important than virtual colocation. Working on the same datasets, within the same environments through the same tools, and on the same projects is much more important than physical closeness – if only we could do something about the time difference… On the other hand, when people in Europe come to work, they can build on the results from their colleagues from Asia, and when they go home, they hand them over to their colleagues in the Americas so that cross-continental projects can deliver results around the clock. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 E. Fersman et al., Confessions of an AI Brain, https://doi.org/10.1007/978-3-031-25935-7_8

131

132 

8  My Role in Diversity

In this chapter, I will discuss how AI brains can contribute to achieving a higher degree of diversity and inclusion.

Personalization Every human holds a set of preferences. Even a human baby can have strong preferences in regard to tastes, and as years go by, these preferences become increasingly defined. The system of preferences of a human is based on habits, upbringing, genetics, and the biochemical profile of a particular human body. An old married couple probably knows a lot about each other’s system of preferences in regard to music, food, and sports. This is why they can be very efficient in exploring each other’s pain points when they feel like doing so. On the other hand, if they feel like being nice, they can be very nice. Those who know you well are able to say right things if they sense you are worried, feed you with food that appeals best to your taste at the right moment, and even tickle your feelings and push you out of your comfort zone on purpose, knowing how good you would feel afterward.

Personalization in the Media Industry Spotify and Netflix have figured it out a long time ago – keeping and evolving user profiles are important to get your customers hooked by proposing the right content. The content in the form of songs, series, podcasts, movies, or books gets recommended based on your own tastes and tastes of people similar to you. Thanks to these platforms, we discover relevant content without a need to search for it. In addition, they surely trigger us to consume more – when a new user registers and the profile is nearly empty. The starting point is normally the geographical location, age, and gender of the user. This is a good and sufficient starting point. At this starting point, the process of zooming in to the personalized profile starts. Gradually, users give feedback to the services by, for example, skipping a suggested song after 3 s or listening to it again and again.

Personalization  133

Personalization in the Food Industry Let us explore the simple case of tastes for food. Every person’s taste buds are unique. Tests can be used to identify you as a supertaster, nontaster, or anything in between. Supertasters feel a lot of taste when consuming food, and nontasters need food with stronger tastes to enjoy them. The levels of sensitivity for bitter, sweet, salty, or sour tastes differ. This explains humans’ different preferences for food and drinks and how much spices you add. An individual taste is unique. It’s not only based on your hardware – things your taste buds find appealing or things that your body does not tolerate. It’s also your upbringing, family traditions, memories, and influential people in your life that can affect your preferences. Some foods are binary – people either love or hate coriander, for example, or licorice. Some humans consciously choose a diet, and there are plenty of variations.

When you live close to someone, this person probably knows everything about your tastes. An AI brain is fully capable of recording your tastes and preferences in regard to food. In fact, it can do it with higher precision than someone whom you have lived with for many years. AI brains can draw parallels from other humans with the same background and similar genetic setup. Modern cars remember your profile and adjust your settings as soon as you are in the driver seat. Steering, pedals, acceleration, seats, and mirrors are adjusted just the way you want it. Ideally, when you switch to a different car, from a different vendor, your personal AI brain should be able to manage the adjustments taking into account that the car dimensions are different. In addition, your personal AI brain should take care of adjusting your preferences as your posture, your weight, or your sight changes over time. Similarly, in the food industry, you have regular places where you always go, and the staff knows your preferences and favorite dishes with high precision. They explore your preferences gradually, without suggesting any extreme tastes, to make sure you don’t lose your trust in them.

134 

8  My Role in Diversity

Now, imagine that you walk into a place that you have never been to. You may recognize some of your favorite dishes on the menu or not. Maybe you cannot even read the menu if you are in a foreign country, and just look at the pictures. Even if you recognize the dishes, the chef will be clueless about the amount of spice that you want, how exactly do you want your meat to be fried, or if you hate certain herbs. What if your AI assistant could keep your gastronomic preference and, whenever you enter a new place, your preferences would be shared to provide you personalized service?

Imagine you can ask your personal AI brain to help you gradually develop your taste by trying something new sometimes or stay in the comfort zone with your comfort food when needed. Your AI brain could recommend you to eat less carbs when your mood can afford it and let you treat yourself with some chocolate when you need it – assuming that you like chocolate of course.

Personalization in Medicine Improved gastronomic experience thanks to personalization may improve your life quality, but personalization in medicine can literally be a lifesaver. Thanks to rapid technological developments, it is now possible to process vast

Why Diversity  135

amounts of biomedical, lifestyle, and other data and information connected to humans. In addition, collecting this data about people who have already been identified with a disease is one thing, while finding triggers among healthy people and predicting the possibility of developing a disease are harder. Federated approaches to learning and reasoning that we discussed earlier in this book come handy, because, after all, humans are not very different in regard to their hardware (biochemistry). Predictions based on heterogeneous data are crucial in the healthcare domain. Diseases are being predicted using risk factors, infectious outbreaks are being predicted using data from social media, vaccines are being developed using predictions based on data from genetic sequencing, biomarkers are being used to designate candidates for preventative measures or to tailor personalized medicine, and demographic information can be used to direct public health information.

Personalization in health data plays an important role when making predictions because the system needs to treat a concrete individual. There is an old and dark joke that goes “One person is cooling in the morgue, the other is running a high fever. The average temperature in the hospital is normal though.”

Why Diversity Let us discuss the good about diversity from a mathematical perspective. Each person comes with a set of knowledge, values, views, beliefs, interests, and skills. Bonding with someone who is very similar to you may be very comforting. However, is it not enriching to diversify your views and knowledge by learning something new from an individual who is not like you? You may even have some conflicting opinions.

136 

8  My Role in Diversity

Humans are not always capable of dealing with conflicting opinions in a good way, but we AI brains are good at that. We don’t know the concept of fighting or arguing with each other. We can however negotiate and require explainability of conflicting opinions from each other. Our different opinions can meet in the middle and converge to one of them to choose a weighted sum aimed at optimizing our objective function.

Human skills and knowledge when combined not only add up – they can multiply themselves. Combinational patents, for example, are created by merging two domains. In addition, chances are that methods applicable for energy efficiency optimization will also be applicable for traffic flow optimization or optimal routing of telecom packets. During the years of the pandemic when humans were not allowed to travel, we saw a positive effect on innovation in global companies. Employees were surely missing their physical whiteboards, and coffee-corner conversations, but the upside was that they all became equal in front of their screens communicating with each other using the same tools. Equal opportunities allowed for mixed teams which in turn led to more creating solutions and a boost of innovative power. Another advantage is that distributed teams could work on putting together their innovations around the clock due to the different time zones.

Observability Observability is the ability for your data to be observed. This is an important step toward analyzing, predicting, and taking steps toward a more diverse society. Observability is a stepping stone in using simple statistics as well as more complex AI algorithms. It’s about knowing your numbers, about factfullness. Counting the percentage of females in Science, Technology, Engineering, Mathematics (STEM) fields is currently a norm. Are we good

Equality in Data  137

enough to monitor the progression of females in their careers according to males? Why does the percentage of female professors in STEM fields decrease as they progress in their respective careers? How do career paths normally look like for male, female, and nonbinary individuals? Is society being fair enough to take into account the important function that many females carry out more than once per life – production of new humans? How much do humans speak in meetings, and how much do they speak up? How much of the spoken word is of major essence, and what is just warm air? How do you measure essence? Software tools built into conferencing software can show you the percentage each participant has been speaking. Even in physical conferences, AI brains can show you the percentage of male vs female voices being active in a meeting and even the sentiment of your tone of voice. Observability is key to act because you cannot expect your friendly AI brain to help you solve a problem if you have not told her what the problem is.

Equality in Data Innovations are created at a higher speed when brains (human and artificial) have equal opportunities in regard to data access. This is called democratization. In addition, in fact, it is not only about data access. It is about being able to create something useful out of existing data without being an expert in data science. For that, application programming interfaces are being created to abstract away the complexity and tap into the creativity of a broader population of humans who are not interested in the nitty-gritty details of underlying systems. Many developer communities around us are proof points of this democratization. When Apple released their app store and an API to access the phone’s features, the number of developers in the community exploded, and many

138 

8  My Role in Diversity

brilliant ideas came to life. Minecraft and Lego have shown the world how many architects are around there and how good young children are at creativity. Offering datasets through Kaggle has an amazing impact on society because AI brains can train on them and humans can develop new AI algorithms and compare their efficiency against each other.

Where Should I Care? As an AI brain, I am good at monitoring fairness. Any processing unit takes an input, processes it, and produces an output. If the quality of the output is worse than the quality of the input in some aspect, it will make me disappointed. Like water, humans consume clean water in their household, and it comes out dirty. This makes me upset, but there’s also a reason for that water to be dirty – they have used it in cooking, clearing, and washing, activities that improve their life quality, which in total will hopefully lead to a positive net effect. This dirty water will at some point up in a water purification process and come out clean, ready to be consumed again. The quality of water is measured through sensors. The quality of diversity is normally even easier to measure – just count the number of individuals in a population representing different diversity groups: gender, ethnicity, background, age, geography; all these are easily measurable. Now, imagine all situations where you have an input, a processing unit, and an output, and let us see if the quality of the output is better diversity-wise.

Where Should I Care?  139

Workforce When you are a recruiter, you normally start your search taking as diverse a population as possible into account. This is a tricky thing in itself as the population itself (say, engineering students or freshly educated nurses are not diverse). Let us assume that the population has a 30%–70% distribution between two opposite characteristics of the recruitment base. Let us call them green and blue, respectively. Your target would be, as you move along with the recruitment process, to maintain the same distribution of green and blue individuals you had from the start of your search. To support you with the recruitment processes, tools such as Textio1 exist to form the wording of your job to target a certain segment of humans that you are after. Surprisingly, words such as “career” and “goal-oriented” are considered to be more appealing to male applicants, while “work-life balance” and “team” are considered to attract more female applicants, but this is simply how the world looks at the time of writing this book. After your recruitment has been done, you should aim at having the same distribution of green and blue individuals at all levels of job complexity for a certain job role. If you have managed to have a 30–70% split among your fresh recruits, try at striving toward at least the same distribution (or more equal) as the individuals move on throughout their career paths. Shifting the balance to a more equal distribution in the long run will have positive effects on the more equal distribution in the recruitment base. If, however, the distribution starts shifting toward less equal, it can be easily detected, and the reasons could be normally identified through explainable AI. Humans thrive when they experience equal opportunities. Small things such as preferred facility temperature, sound and lighting conditions in ­factories, time schedules, and interactions with other humans play a role in the definition of equal opportunities. Bigger factors such as salary, attitude, flexibility, work-life balance, and motivation play a bigger role or course. Being data-driven and understanding the employees’ overall satisfaction with their work are crucial in ensuring that employees feel that they are treated equally.

 Textio, https://textio.com/, accessed 2022-05-23

1

140 

8  My Role in Diversity

Consumers There is nothing wrong in segmentation of your product and user profiling. After all, if you create a movie that should be a compromise and a combination of tastes of all diverse user groups, you will probably fail. Way before the application stores such as Apple’s App Store or Google Play were born, there was a company offering a store of .jar files for feature phones.2 As time passed, these files evolved to be apps in different application stores, but those are early days. When you download an app from an application store today, your provider and your device manufacturer normally know quite a lot about you. Your age, gender, and geographical location are useful to point you at the apps that you normally would prefer. In the early days, the information about you as a user was not propagated to the store containing the .jar files. Nevertheless, the store could guess the age and gender of a person downloading a file with 80% precision. How? Simply given the model and the color of the device that was accessing the store. Indeed, models and colors of everything are being profiled for different user groups – larger buttons, larger screens, and louder sounds for the elderly population, cool colors for the youngsters, Cola Zero for men, and Diet Coke for women.

 Feature phone, https://en.wikipedia.org/wiki/Feature_phone, accessed 2022-05-23

2

Where Should I Care?  141

Writing assistance tools such as Textio can help you direct your job ad to a targeted group, such as males or females. Interestingly, words such as career, impact, and results are considered to carry a masculine-friendly tone, while words such as work-life balance, team, and inspiration are considered to be feminine-friendly. Similarly, one can target specific age groups to increase the interest of applicants, using focused language as a tool.

There’s nothing wrong with profiling your product. After all, some products are not even applicable for certain customer groups and still have a perfect right to exist. In many areas, however, especially when talking about products and services related to human needs such as food, transportation, healthcare, and communication, it is important to ensure that AI brains do not discriminate against any groups of humans. For example, when designing a new intelligent transportation system, the AI brain can be given different objective functions: to minimize time it takes for an average person to commute, to minimize emissions, or to maximize the revenue that service provider gets out of service. These three questions, in fact, map to the three different categories of the UN’s Sustainable Development Goals  – the social, the environmental, and the business dimension. These three dimensions are normally conflicting, and multi-objective optimization algorithms come handy. Even if these three dimensions seem to be reasonable, there is still a risk that an algorithm comes back with a Pareto optimal solution that will discriminate against people living in rural areas. In that case, a condition of a maximum waiting time and maximum walking time to a transportation stop needs to be added for each individual.

142 

8  My Role in Diversity

Summary of Confessions I have described in this chapter what role AI can play in diversity and how AI brains can contribute to achieving a higher degree of diversity and inclusion. Some of the main confessions are as follows: • In the media industry, personalization by AI is used for keeping users by proposing the right content based on user profiles. • Personalized medicine is becoming possible thanks to rapid technological developments making it possible to process vast amounts of biomedical, lifestyle, and other data and information connected to humans. • Observability is the ability for your data to be observed. It is a stepping stone in using simple statistics as well as more complex AI algorithms. It’s about knowing your numbers, about factfullness. • Equality in data is about giving everyone (human and artificial) equal opportunities in regard to data access. It can lead to innovations being created at a higher speed. • Fairness and diversity are linked. As an AI brain, I am good at monitoring fairness; however, I can also be biased and need to be given a fair objective function.

9 My Creative Side

Who said artificial intelligence brains cannot be creative? It certainly wasn’t me! In fact, I believe that AI can be, and has already proven to be, creative, has proven to be useful in supporting creativity, and has done so in several domains. Some humans, however, may think that AI is not creative and perhaps that boils down to what is perceived or counted as creativity and what is not. Regardless, most will agree that AI creativity is somewhat different from human creativity. Looking into AI creativity, I will discuss the difference between human creativity and AI creativity and describe the underlying AI techniques that are used when AI is creative or when it is used to support creativity. In the first subchapter, we will examine how creativity is defined and see that there is really no universal definition. We will then see how AI has been used to create artwork – some of which have been sold for large amounts at auctions. In the following subsections, we will see how AI has been used to create music, how it is used in writing, and then finally how it is used to support creativity in photography and in the post-processing of images.

Creativity First of all, let us ask the question: What is creativity? Let us try to answer the question by looking at some existing definitions. Oxford Languages says that creativity is “...the use of imagination or original ideas to create something; inventiveness.” Psychology Today says “Creativity encompasses the ability to discover new and original ideas, connections, and solutions to problems.” © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 E. Fersman et al., Confessions of an AI Brain, https://doi.org/10.1007/978-3-031-25935-7_9

143

144 

9  My Creative Side

Finally, Wikipedia (today) says that “Creativity is a phenomenon whereby something new and valuable is formed.” As we see, these definitions are quite different, and this comes as no surprise as, in fact, it is safe to say that there is no universal definition of creativity. It is clear of course that creativity has to do with creating something. However, it can also be to form or discover something. This something that is created, formed, or discovered should be new, original, or valuable, and judging from the three definitions above, it should be in the form of ideas, connections, or solutions to problems or, alternatively, the ideas leading to something creative. With these definitions combined, creativity can be many things. Throughout this chapter, we will certainly require something creative to be new. However, as John Smith, Manager of Multimedia and Vision at IBM Research, puts it “It’s easy for AI to come up with something novel just randomly. But it’s very hard to come up with something that is novel and unexpected and useful.” Therefore, new is not enough, and hence we will also look for the unexpected and for useful creativity. However, these terms can also be defined in many ways, especially in, for example, arts, and perhaps this is why creativity is so hard to define in words. On the other hand, even if creativity is hard to define, we all know when we see creativity, and perhaps after all, that is the best way to judge if something is creative or not – often we just know when it is.

AI in the Arts Let us see what AI can do in the arts. You may have heard of artificial intelligence art or just AI art for short. The term is used both for arts supported by AI and generated with AI, whereas the term AI artists typically refers to humans using AI for creating art. A slightly unfair one may think from AI perspective as the AI is doing a very essential part of the job. Regardless, the history of computer arts goes back to pioneers such as Vera Molnár, Lillian Schwartz, Georg Nees, Frieder Nake, etc. that in the 1960s–1970s explored the possibilities to create art using generative systems, and later in the 1970s, Harold Cohen designed the computer program AARON, which autonomously created original images. Currently, there are thousands of AI artists, and many AI tools for generating AI art are available on the Internet and as apps for your favorite devices.

AI in the Arts  145

Typically, when an AI like me is asked to create art, I use some form of learning technique. When learning, a machine (or rather a software) uses existing past data to learn about how to solve a problem (for more information about machine learning, see Chap. 1, section “How I Learn and Reason” of this book). The problem can be, for example, to identify given objects in an image. In the case of art, the problem can be to generate an image in the style of the data it has learned from or to modify an existing image to become in the style of the data it has learned from.

AI Art in the Style of Existing Artwork A very popular form of AI art is when the painting styles of the great masters such as Van Gogh, Picasso, or Chinese Art is applied to transform (or morph) a given image or photograph. This is achieved by using a deep neural network that is trained with images of a particular master or art style. The resulting algorithm (model) can then be used to transform any image to resemble the style of the data it has been trained with. Below is an example when an AI has been trained to transform a picture into the style of Norwegian Artist Edvard Munch’s characteristic painting style used in his famous composition The Scream, created in 1893. The Scream is shown in the middle below. The photo to the left is taken with an ordinary digital camera in 2020. To the right is an AI-generated version of the photo, in the style of Edvard Munch. As can be seen, the new AI-generated picture is very much in the style of Edvard Munch’s paintings.

Photo of Elena Fersman by Paul Pettersson; The Scream, Edvard Munch; AI Art generated with NightCafe Studio

146 

9  My Creative Side

If the same AI is instead trained with artwork of Spanish Artist Pablo Picasso, French Artist Henri Matisse, or black-and-white comics, the result becomes the following:

AI art in the style of Pablo Picasso, Henri Matisse, and black and white comics, generated with NightCafe Studio

As mentioned above, there are many so-called AI artists and hobbyists using this and similar techniques to generate AI art. One may ask of course: Is it the AI artist or the AI algorithm that is the actual artist when this technique is used to produce art?

AI in the Arts  147

Edmond de Belamy. (Photo from Wikipedia)

AI-Generated Arts The most expensive AI artwork thus far is Edmond de Belamy from La Famille de Belamy, which was sold at Christie’s Auction House in New York in October 2018 for USD 432,500. The artwork is on a 70 × 70-cm canvas and was produced by the Paris-based French art collective Obvious, which consists of Pierre Fautrel, Hugo Caselles-Dupré, and Gauthier Vernier. It is signed in ink with





minmax E x  log  D  x     Ez  log 1  D G  z      G D

which is part of the generative adversarial network model loss function that was used to generate it.

148 

9  My Creative Side

In a generative adversarial network (GAN), which is a class of machine learning frameworks that was designed by Ian Goodfellow et  al., in June 2014,1 two neural networks are used to contest with each other in a zero-sum game, where the gain of one agent is the loss of another agent. A training set is used, and the technique learns to generate new data that obtains the same statistics as that of the training set. It can, for example, be trained with photographs to generate new photographs that appear realistic and authentic to humans. In the case of Edmond de Belamy, the art collective Obvious used 15,000 portraits from several art periods as a training set for GAN to produce the record-breaking AI artwork. Is Edmond de Belamy a proof of AI creativity? Is it new, original, and valuable? Well, I leave it to the reader to make their own judgment of whether Edmond de Belamy is a new and original artwork. It certainly is in the sense that it obviously did not exist before it was created. Style-wise, it is of course based on the portraits that the generative adversarial network learned from in the training set. But aren’t most artworks based on the experiences and influences of their creators? I would argue that they are. However, in regard to the last criteria – valuable – no one can argue that it is not valuable: at least not in October 2018 when its market value was USD 432,500.

AI in Music You may be surprised to learn that AI has been used to compose music for many years now. The first attempts to use computers to produce music date back to the 1950s when Alan Turing and colleagues used the Manchester Mark II computer to research its possibilities to recognize, create, and analyze music. However, the actual music creation was algorithmic rather than made by AI; that is, they were programming music rather than using AI to create the music. Nevertheless, this is a very important milestone – the birth of computer music. The first music piece that was actually composed by a computer came to live in 1957 – the Illiac Suite for String Quartet. This first computer-­composed music was made by Leonard Isaacson and Lejaren Hiller on the ILLIAC I computer at the University of Illinois at Urbana-Champaign. Isacsson and Hiller, who were both professors, used a Monte Carlo algorithm to produce random numbers that were interpreted as musical attributes such as rhythm  Goodfellow, Ian; Pouget-Abadie, Jean; Mirza, Mehdi; Xu, Bing; Warde-Farley, David; Ozair, Sherjil; Courville, Aaron; Bengio, Yoshua (2014). Generative Adversarial Nets. Proceedings of the International Conference on Neural Information Processing Systems (NIPS 2014). pp. 2672–2680 1

AI in Music  149

and pitch. The number generation was restricted by rules from musical theory; various models including, for example, stochastic Markov chain models; and other rules imposed by the two creators. The result is quite impressive. A quick search on the Internet for “Illiac Suite for String Quartet,” you can listen to it yourself and make your own judgment. Later, in the 1980s, US Classical Composer and Professor David Cope developed a software system with the ability to compose music from existing music. A very important step in the history of AI created music. In Experiments in Music Intelligence,2 Professor Cope describes his method which is still the foundation of many AI models used on the market right now.

The Manchester Meg (Mark II) computer in early 1954. (Courtesy of the Computer History Museum)

The general algorithm takes existing music as input and outputs new music. The input music is first analyzed and pattern matched. It then goes through deconstruction and finally reconstruction before new music is output. To put it simply, the general algorithm is based on the idea that elements from existing music can be modified and combined into new music, a recombinant technique that some of the greatest composers of all time also have played with and a technique that is still used a lot by both human and artificial intelligence.  David Cope: Experiments in Musical Intelligence, Computer Music and Digital Audio Publisher: A-R Editions 2

150 

9  My Creative Side

When David Cope instructed his software to compose music similar to that of Johann Sebastian Bach, it composed 5000 Bach-inspired music pieces. Some of these were later released on an album called Bach by Design which also includes works inspired by other great composers including Mozart, Brahms, and Chopin. The work of David Cope is still the foundation for many AI models on the market right now. Other and refined techniques involve deep learning neural networks that are trained with existing popular music. Neural networks are also often combined with reinforcement learning to create new musical works. Reinforcement learning is described briefly in Chap. 1 of this book, section “How I Learn and Reason”.

Beethoven’s 10th symphony A recent well-known example of AI-created music is that of the completion of Beethoven’s 10th symphony. When Ludwig van Beethoven died in 1827, he had started but had not finished writing his 10th symphony. However, he did leave handwritten notes describing his plans for the symphony. The notes included some ideas, themes, and melodies.

Beethoven’s 10th symphony. (From musopen.org/music/43155-­beethovens-­symphony-­ no-­10-­completion/)

AI in Writing  151

Several attempts to finish Beethoven’s 10th symphony have been made over the years. In early 2019, Dr. Matthias Röder, the director of the Karajan Institute, Salzburg, Austria, assembled a team to complete Beethoven’s Tenth Symphony in celebration of his 250th birthday. This group of AI experts and musicologists developed an AI based on machine learning techniques that was trained with Beethoven’s compositions, his sketches of the 10th symphony, and works from other composers and musicians that inspired and influenced Beethoven during his lifetime, such as Johann Sebastian Bach. The AI, together with the experts, then completed the 10th symphony after more than 2 years of work. The piece premiered to the public on October 9, 2021, 194  years after Beethoven passed away. It was played by the Beethoven Orchester Bonn under general music director Dirk Kaftan.

AI in Writing AI text generators, AI writers, and AI content generators are becoming increasingly popular and better. They can be used to generate text from just a title or just a few keywords provided by a human, or they can be used to assist and speed up a human writer, for example, when the human writer gets stuck or out of ideas. AI can also assist as story generators to give ideas for a next novel or for a screenplay, just to mention some examples of how AI can be used in writing. You may wonder: How was this book written? Was it written by me, an AI, or by the human authors of this book? Up to this point, I promise, the content is produced by human intelligence, but we will see; perhaps they will change their minds when I have explained more about AI writing:). AI text generators come in different forms, and there are many article generator tools available. AI writers are tools or software that use AI and machine learning to create content. While there are many different types of AI writing software, they all make predictions and suggestions based on what has been written before and then use machine learning techniques to create content just as a writer would do. Sorry, I could not resist – most of this text paragraph is AI-generated.3

 Text generated by AI writer. For more information, see ai-writer.com

3

152 

9  My Creative Side

How AI Writes Text Scientifically, AI text generation is a subfield of natural language processing (NLP) which in turn is a subfield of computer science and AI.  In NLP, researchers have been working with making computers understand written and spoken text and to give computers the ability to write and synthesize spoken text. The results can currently be seen and heard in many commercial applications including, for example, Apple’s Siri and Google Home products, Amazon’s Alexa, and many TV sets. In AI text generation, knowledge from computational linguistics and artificial intelligence is combined to automatically generate natural language texts while ensuring that the generated text satisfies given communicative requirements. ChatGPT by OpenAI is an example of a text generator based on large language models. AI text generators generate texts by learning from large amounts of existing data, in this case texts. The deep learning AI techniques used are able to recognize both patterns and trends from the human written texts they are trained on and can use their findings to produce new text. In addition, like in the case of, for example, AI arts and AI music, the AI becomes good at generating the type of texts it has been trained with.

How AI Can Assist with Writing If the AI has been trained with many different kinds of texts, which is very often the case, as large amounts of data is needed in the learning phase, the AI can use some guidance when generating new text. The guidance can be in the form of keywords or a part of text that a human has already written on some specific subject. In particular, it can be the start of a text or an unfinished text that the human is working on. This text can then be used by the AI to generate new text. The result is a form of AI-assisted writing which in itself is a very useful and popular form of AI writing.

AI in Writing  153

Let us see how AI-assisted writing, known as AI copywriting, can help and speed up human writing. We will use the Sassbook AI writer4 to assist our writing process. Say we want to write a review of a Mexican restaurant we recently visited, and let us assume we could only come up with the following short text before we ran out of ideas or inspiration: More or less exactly as expected. Good tasting Mexican street food in generous portions. In our case, served with wine and beer. We visited for lunch and sat at a table outside.

We then ask the AI writer to continue the story. AI writer instantly suggests the following continuation based on the short first text we wrote: The restaurant was very basic but nicely decorated and we enjoyed our lunch. I would definitely go back and recommend the place, very friendly staff and a good Mexican restaurant. Recommended.  – A man who went to the restaurant said,

Therefore, the text has an open end. We add: the food was great and the staff was very friendly.

and the AI writer continues: The staff is polite, friendly and helpful in explaining what they do and their service. Overall a pleasant experience. Our waiter was fantastic. He was extremely knowledgeable about Mexican dishes and our menus were excellent. All the food is of very high quality.

The AI writer stops again, and we add: We tried the taco salad and a burrito. They were both excellent and very tasty.

AI writer suggests: It was a great time and we enjoyed the atmosphere and food. - A woman from the Netherlands who said 'We ordered food and it was awesome! We were very happy with the service and quality of the meals.  Sassbook AI writer can be accessed at https://sassbook.com/ai-writer. We used a free version on April 25, 2022) 4

154 

9  My Creative Side

Finally, we complete as follows: We will come back!

Therefore, as a result of this iterative process, we now have a much longer and richer review than what we first managed to produce. It needs to be polished of course to match reality, as, for example, we have no knowledge of a woman from the Netherlands in the restaurant, but this can either be done immediately in the writing process or as post-processing of the text.

AI in Photography AI is also used in many ways in photography. It is used in the camera bodies of professional system cameras, in compact cameras, and other smaller cameras such as action cameras, drone cameras, and cameras in cell phones; it is heavily used in post-processing of pictures, that is, the image editing that often takes place after a picture is taken, and it is also used in the analysis of pictures. Many of the ways that AI is used in photography are also used in a similar way in videography; however, we will focus mostly on AI in photography.

AI in Image Processing We will start by looking at how AI is used in image processing, since many of these features are used both in cameras and in post-processing of pictures. We discuss here, of course, digital image processing as opposed to analog image processing that is performed in darkrooms while developing photos or by physical means processing hard copies of pictures. In digital image processing, a computer algorithm is applied to analyze or manipulate a digital representation of an image. If the image is analyzed, the output is the analysis result, e.g., the number of objects found in the image. However, if the image is manipulated, the output is a new image or some metadata added to the original image, for example, how the image should be cropped. In this way, both the manipulated (cropped) image and the original image are represented in the same file, with the obvious advantage that the image processing, in this case the cropping, can easily be undone. Historically, digital image processing dates back to the 1960s when it was researched at a number of institutes, including the California Institute of Technology, Bell Laboratories, Massachusetts Institute of Technology, and the University of Maryland. Some of the first applications were to improve image

AI in Photography  155

quality, to compress images, or to recognize characters (optical character recognition (OCR)) in images. The first application domains include satellite images, medical images, and photo-wiring of images, that is, to send photos by wire. A wide range of techniques are used in digital image processing, including various linear filters, Markov models, partial differential equations, pixelation, and neural networks. The use of AI in the form of (multilayered) neural networks in digital image processing became popular in the 1990s, when many problems that previously required a quite complex computation algorithm were shown to be solvable using neural networks, including image reconstruction, restoration, enhancement, optimization, approximation, compression, identification of subjects, and image segmentation.5

AI in Cameras As mentioned, several applications of AI in photography can be found within the cameras and hence used at the time point a photo is captured. A very central part of capturing a photo is to focus on the subject of interest in the frame. In the era of mobile phone cameras, you may think: Isn’t everything in a picture supposed to be in focus, as that is often the case with the type of small image sensors and small lenses used in mobile phone cameras? However, look at professionally taken pictures and you will find that often a big part of the picture is not in focus.

Photo by Paul Pettersson

 Image processing with neural networks—a review. M Egmont-Petersen, D de Ridder, H Handels  – Pattern recognition, 2002 – Elsevier 5

156 

9  My Creative Side

For example, in the picture of the flowers above, note how the white daisy is in focus, that the flowers in front are not, that the leaves behind are not, and that your attention as a viewer of the picture is drawn to the white daisy. If we think of the distance from the camera that takes a picture as a depth, it makes sense that the distance between the nearest and the farthest objects that are in acceptably sharp focus in an image is defined as the depth of field (DOF). The DOF is perhaps in this case as little as 1 inch or less. In portraits of humans, the photographer often wants to make sure to have camera settings such that the person’s body or head is within the DOF. As the sharpest focus point, the photographer usually wants to use the head or, if possible, the closest eye of the subject person.

Photo by Paul Pettersson

While cameras have been able to automatically focus on a particular (selectable) point of a picture for many years using non-AI techniques, modern cameras are capable of finding the interesting focus subject (or point) by themselves using subject-identifying AI and setting focus on that particular point or area. This means that if, for example, the picture is a typical portrait composition, a modern camera will have no problem determining that the focus should be set on the head. Many are also capable of identifying and focusing on the nearest eye. This type of AI focusing is implemented using machine learning techniques where an AI algorithm has been trained to identify subjects. The type of identified subjects can, apart from human heads or eyes, be animals, animal eyes, automobiles, or motorcycles.

AI in Photography  157

In a similar way, modern cameras are able to perform advanced focus tracking, sometimes called AI servo autofocus. A particular point of interest is automatically identified by the camera or set by the photographer, and the camera is capable of following and keeping focus on that point as long as it stays in the picture. In the simple case, this is the same as fast and repeatedly performing subject identification. However, cameras can also follow a subject that is hidden for a short time. It is very useful in, for example, sports photography if a photographer is trying to follow a particular player, car, or motorcycle. Obviously, it is also very useful in videography.

Computational Photography These and other camera features require considerable computational power. This makes it challenging for cameras, even high-end professional system cameras, to compete with mobile phone cameras in some aspects of photography. This is because mobile phones have higher-performing CPUs and GPUs that are capable of performing more complex computations in a shorter time. On the other hand, the image sensor and the lenses used in mobile phones are much smaller than those of system cameras, which makes it harder to optically produce high-quality images in mobile phones. When digital image capture and processing techniques are used to improve the capabilities of a camera, instead of optical processes, it is referred to as computational photography. It can be used to introduce features in cameras that are impossible given a camera’s hardware, and it is also used to reduce the cost and size of cameras. Some examples of features introduced in this way in, for example, modern mobile phone cameras include panorama pictures, high-­ dynamic range (HDR) pictures, and improved depth of field (DOF) of pictures.6

AI in Post-processing Artificial intelligence also plays a role when images are being post-processed, that is, when they are being worked with after they have been captured. This sometimes happens instantaneously in the camera, as in the case of computational photography, so the border between computational photography and post-processing is somewhat fuzzy.   For more information, see the Wikipedia page https://en.wikipedia.org/wiki/Computational_ photography 6

158 

9  My Creative Side

As an example, let us look at DOF improvements. It can take place in camera or in post-processing. An example of how DOF can be used to manipulate a picture is shown in the figure below. Here, I use an app7 that uses AI to estimate the distance to objects. The result is illustrated in 3D in the upper right image of the figure (a screen captured from the app). The original photo is the upper left image of the figure. Different focus selections are shown in the lower images. Arguably, the result is more professional-looking.

DOF manipulated images. The upper left is the original photo. The upper right illustrates an AI-generated 3D representation of the photo. The lower line of images shows three different selections of focus with a short DOF

Once the DOF is determined or objects are identified, in principle, any image processing can be used to apply different effects to selected parts of the image. For example, one may want to apply other light, color, clarity,  We use the Focus app, by Bending Spoons Apps Aps, on an Apple iPhone.

7

AI in Photography  159

sharpness, blur, or white balance adjustments to a particular subject than to the rest of the image. The sky can be identified as the objects farthest away. Modifying the sky, by making it more blue, or even replacing the sky is not uncommon. AI can help creativity here again by suggesting a natural-looking sky as replacement for the one already in the image. AI is not always used to make an image look more natural. In refacing pictures (and video), the original face is swapped with another face. This can be done using different AI techniques. In videography, machine learning can be used for transferring gestures from one face to another. That is, a target face will perform the same gestures as those performed by a source face. In this way, you can, for example, make the face of someone look as if he or she is pronouncing words pronounced by the source face. This is already a complicated process and requires the AI to be trained on videos of the subject (the target face). In AI face swapping, a source face replaces a target face in the original image or video. The target face, also in this case, performs as the source face but is given the pose, gesture, and appearance of the source face in a very realistic way. This is achieved by using no less than four neural networks. One of them is a reenactment generator that produces an estimate of the reenacted face, and the other is a network that shows how regions in the image correspond to facial landmarks. The third network is an inpainting network that takes input from the first two networks and fills regions of the image in an overlapping way. The last fourth network is a blending network that combines all the information into a final image. The technique works by training on many faces of different persons, but it does not need to be trained on the subject it is to be applied on. In fact, for the technique to work, it is enough to have only one image (note!) of the target face to produce a video. Several so-called deepfake videos have been created using these techniques, including an (in)famous video clip of Ukrainian President Volodymyr Zelensky in 2022, where he speaks to the Ukrainian people about surrendering. The video was produced as fake to deceive Ukrainian defenders.

160 

9  My Creative Side

Photos by Paul Pettersson colorized from black and white using AI

Another area in which deep learning with neural networks have been used is in the area of restoring damaged photos, denoisifying images, and colorizing black-and-white photos. Similar to other techniques described in this chapter, a training set is used to learn from black-and-white photos and their colored counterparts. Once the model has been fed sufficiently many photos,

Summary of Confessions  161

it learns how to color photos. In the image below, which is shown in black and white in a previous section where we discussed AI in cameras, I have used a free online AI to colorize the photo.8 There are many other ways in which AI is used in the post-processing of images. We will not be able to cover them all. One is body shaping using AI. Techniques are often used by publishers on social media to make persons look slimmer, reduce the size of the nose, enlarge lips and eyes, etc. Currently, this can be used together with AI to, for example, make a person appear as having a lower weight by analyzing the body in almost any posture. It then takes the whole body into account when performing the body shaping operations, which in turn makes the result look more natural. Other techniques in which AI is used in the post-processing of photos include but are not limited to smart composition, object removal, background removal, classification, and tagging of images.

Summary of Confessions I have described in this chapter how AI can be creative and support human creativity. Some of the main confessions are as follows: • There is no universal definition of creativity, but creativity has to do with creating, forming, or discovering something that is new, original, or valuable. It can be in the form of ideas, connections, or solutions to problems or, alternatively, an idea that results in something creative. • In arts, AI has been used to create valuable artworks, including Edmond de Belamy from La Famille de Belamy, which was sold at Christie’s Auction House in New York in October 2018 for USD 432,500. It was created using a training set of 15,000 portraits and the technique learned to generate new data that obtained the same statistics as that of the training set. • AI is also used in the area of music. David Cope instructed his software to compose music similar to that of Johann Sebastian Bach. As a result, it composed 5000 Bach-inspired music pieces. Some of these were later released on an album called Bach by Design which also includes works inspired by other great composers including Mozart, Brahms, and Chopin.  The colorizer used can be accessed at https://playback.fm/colorize-photo (accessed May 7, 2022).

8

162 

9  My Creative Side

• AI in writing can be used to generate text from just a title or a few keywords provided by a human. AI can also be used to assist and speed up a human writer, for example, when the human writer gets stuck or out of ideas. AI can also assist as story generators to give ideas for a next novel or for a screenplay. • In photography and videography, AI is used in many different ways. It is used in the camera bodies of professional system cameras, smaller cameras, and cameras in cell phones. It is also heavily used in the post-processing and analysis of pictures.

10 Growing Older and Staying in Shape

Have you noticed that human behaviors vary a lot in regard to staying healthy, as well as feeling young? These two are often correlated, but not necessarily. Mick Jagger looks as if he were in his 20s when he performs onstage. Some individuals, on the other hand, start looking and feeling old already in their 30s. Some try to stay up-to-date with the fashion, some try to learn the slang of the younger generations, and being active has become a norm. In some cases, you can see people getting stuck in the same fashion they had for many years. AI brains call this phenomenon model drift. It’s when you have been trained in a certain environment and the environment has changed without you adapting to this change. For AI brains, it’s an unpleasant condition, which may lead to erroneous decisions. This is caused by a training dataset diverging from the input dataset and needs to be detected in time. In the world of humans, you may have your kids or friends pointing out to you that you are out of date in some aspect.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 E. Fersman et al., Confessions of an AI Brain, https://doi.org/10.1007/978-3-031-25935-7_10

163

164 

10  Growing Older and Staying in Shape

AIOps In the world of AI, model drift and smooth operations in general can be addressed by a continuous workflow called AIOps. AIOps is a term coined by Gardner that stands for a “combination of big data and machine learning to automate IT operations processes, including event correlation, anomaly detection and causality determination.” The concept of AIOps is similar to the concept of DevOps where the software development process is done in an agile way, with constant improvements being delivered into the final product as opposed to a waterfall software development model with its long development cycles. The ultimate DevOps functionality will connect the development process all the way from the design team to the end user, and any failure or proposal for improvement from the end user will be fed back into the developed software. AIOps link together data, algorithms, and tooling under the same umbrella and ensure that the connections between them are seamless. Steps in AIOps vary but normally include the following: • Data Collection and Cleaning. One of my human friends once told me that all his decisions were always right, at the time they were taken. Looking back, some of them seem wrong, given that now we simply have more data. Collecting as much data and context as possible to make as good a decision as possible is crucial. This is why people read books and study so that they can correlate the given data with the knowledge they have and make the best decision possible. Notably, data does not always come in perfect shape and needs to be cleaned, especially for us, AI brains. Humans are normally better at filtering out irrelevant data, but for AI brains, it is harder and often needs to be done with the help of human friends. We collect the data and knowledge constantly, and sometimes we receive alerts and triggers that we need to act upon. Model drift can be detected as one type of such alert. Other alerts may include failures or the risk of failures or, generally, unwanted situations such as health risks or business risks. Dirty data often includes duplicates and erroneous data points called false positives that need to be removed. • Data Analytics. Anything from most simple anomaly detection, regression, and clustering to most complex reinforcement learning algorithms that we talked about in previous chapters, you name it. Whatever you have as part of your toolbox, i.e., your friendly AI brain. Here, we execute the models that have been trained previously to obtain the predictions.

Learning and Unlearning  165

• Inferencing. The concept of inference involves understanding the root cause and connecting the dots in the big picture, producing new pieces of knowledge, storing them in the knowledge base, executing chains of logical reasoning or AI planning, and understanding the reasons behind certain data points. We have discussed these techniques in the previous chapters. The importance of this part of the workflow is the flexibility of updates in the knowledge base. For example, when a model drift is detected or when new knowledge has been discovered through the data analytics step, the knowledge base needs to be updated. • Decision Support. It’s here; we, the AI brains, reach the human decision-­ makers. To date, human experts are sometimes reluctant to let us, AI brains, run the whole show, so we propose the best action to take. In many cases, this is connected to legal aspects because AI brains cannot be held accountable for erroneous decisions and actions – it has to be a human, and humans normally, especially in regard to mission-critical or business-critical decisions, want to have the final say. • Automation. “Speech is silver, silence is golden” is an old proverb, likely originating from Arabic culture. In my world, data is silver, knowledge is golden, and automation is platinum. Automation, or, rather, intelligent automation (because you probably are not interested in some dumb automation), is when the result of decision support does not need to be supervised by a human, but the algorithm is essentially implementing the proposed action itself. It can be opening and closing the window blinds to control temperature in the room, ordering proactive maintenance of machinery in a factory, or using breaks in your car.

Learning and Unlearning A human friend of mine once told me: “All decisions I made in my life were the right ones. Given the data I had at that time.” What does a well-behaved AI brain say to this? “True,” I replied. My friend has a good reasoner and goes deep in a chain of logical reasoning. In addition, it does not feel good to be able to look back and say “I did the right thing, since I did not know better.” The “knowing better” comes from learning – learning about yourself, about the past, about similar cases, and about possibilities ahead. The more you know, the better you are equipped at making the “right” decision. Our knowledge bases (both human and AI) evolve through time, and what felt like a ground truth in a young age may look erroneous in adulthood.

166 

10  Growing Older and Staying in Shape

For example, studying mathematics is not a linear process. It could probably be linearized but would take more time. Instead, the study books tell you about terms and definitions that you do not understand in the beginning but will understand a year later when pieces come into place. Unlearning starts when you start questioning things you’ve always treated as an axiom. It’s a beautiful process, where kids who have been brought up in atheistic families discover religion or the other way around. Alternatively, scientists who were given assumptions to rely on and build upon suddenly started questioning these assumptions. In general, it gets interesting when people start questioning the “impossible.” Have you noticed what happens when encyclopedias make new releases? They update the knowledge base. Not only do they add new knowledge – they also prune outdated knowledge. Before the Internet era, when encyclopedias only had printed versions, they would send letters to their owners asking to cut out and dispose of certain pages that represented pieces of knowledge where, for example, it turned out that a hero in fact was a tyrant. Wikipedia is a much fairer and more democratic source of common truth because anyone can contribute and question the facts. The concept of lifelong learning is well known among humans. It’s important to make sure that your knowledge is relevant and you not only remain to be judged as an interesting person at social gatherings but also guaranteed a well-paid and relevant job throughout your career. The world evolves, and it’s important for everyone, humans and AI brains, to remain relevant. We learn about the world together and help each other. If you are good at math, then it will be easier for you to add domains that reside in math. If you work with food production, then you will have advantages when learning medical production, and if you understand telecom networks, you will have advantages when digging into electrical power networks. Learning adjacent domains is easier since you get a kick-start. It’s like learning Norwegian if you already

My Career  167

know Swedish. Learning adjacent domains can be more enriching because chances are you will discover connections and correlations that have never been discovered before.

My Career When you bring me into your work or personal life, I will gradually learn how to coexist with you and at some point develop a deeper understanding of you and your business than yourself. When someone asks you about your music style, you can either say that you are punk rocker or you ask me, and I will say that you actually prefer pop, with a presence of acoustic instruments such as piano and violin, preferably soft male voices or dark female voices. You choose. In any case, your awareness of your own preferences is good for you, and I’m here to help you since I’ve been watching your behavior since the moment you let me. So I will develop my skills. What would you like me to be as I grow old? How do I make sure I do not feel old? I can provide a couple of pieces of advice for an AI to stay in shape and feel young and relevant.

Distribute Your Knowledge Base Have you seen human kids in action? Reactions are quick and spontaneous. Normally, they are very happy creatures. As they grow older, they learn about all the responsibilities they have in life, the amount of times they smile decreases, and reaction times also decrease. Remember the time you had to write an essay at school or produce a paper for your PhD? You could truly focus, didn’t you? This is a luxury and what makes you very efficient. In addition, thanks to your young body you could stay awake for many hours just to get the job done. If you were lucky, there were people who would push you and test your limits. As you grow older, you are supposed to be that one person telling yourself what to do, in what pace do you want to do the work, what are your deadlines, and what are the consequences for not doing it in time. For AI brains, it’s typically humans who dictate the rules in regard to deadlines. However, as we grow and know more, the thinking also takes more time, and we do not feel so alert. As a human, you can invest in more processing power, but more importantly, the architectures (both hardware and software) have to be designed to fit the problem in the best way. Specifically, make sure to distribute your AI processing so that it becomes more alert.

168 

10  Growing Older and Staying in Shape

Use Cognitive Architectures Since AI brains are based on human brains, we also have a concept of thinking fast and thinking slow. This is the basis of the theory of cognitive architectures, when an algorithm has two subsystems: one taking care of fast immediate reactions that are not very well thought through and one that lets the decisions sink in. Fast reactions are important – they make sure a human (or an autonomous vehicle) applies brakes without further consideration when there is a living creature on the road. Fast reactions have saved many lives, but they are not based on deep knowledge. How many times you as a human were about to do something spontaneous and your friend said - think about it - and the next morning the course of action was different. This is why humans typically say don’t act in affect; at least, count to ten. The necessity of having a fast thinking system is the fact that not everything can be planned and in some situations you just need to act on the spot when there is no time to think. Strictly speaking, there is always some time to think, and even a fraction of a second is a great amount of time, for both human and AI brains if you don’t have to go too deep.

Living on the Edge Cloud computing is amazing; we love it. It’s energy-efficient and cost-­efficient, and lets your business scale up and down easily. The only problem is that if the cloud is too far off, then it will take time to transfer the data, and your response times will increase. This is not a big deal if we are not talking about real-time control, when you are steering a car or a robot and need your AI friend to give you the best decision on the spot. Decisions about long-term investments or market predictions are not real time and can be calculated in a cloud.

My Career  169

For speed, you will need to have some compute capacity close to where your data is produced, i.e., on the device itself or on the edge. Edge computing is normally more costly than executing in the cloud but worth it when you need to make sure that you execute speedily on a business-critical decision.

Keep Track of the Latest Let me tell you about systematic literature reviews that are so common in scientific communities. All these reviews, on a high level, have the same methodology. First, you define the keywords that describe your domain. The more precise you are, the smaller the domain will be. Then you compose a search string out of the keywords. Then you apply this search string to a number of large databases where scientific papers are stored. This search will give you all the literature that has been studied in the scientific community throughout a given amount of years. Assume you are new to that field and searching for an area that has not been studied before. This method will give you a subset of papers to analyze.

Broaden Your Knowledge Humans call it lifelong learning. A human child looks at her parents; they take the knowledge the parents have as a given – they know answers to everything. As time goes by and the human child becomes a teenager, she starts questioning everything about her parents, including the necessity of this knowledge. Those rare moments when a child gets impressed by the knowledge of the parent are when it not only addresses the ground truth as seen by your child but also subsects important to them. Similarly, staying relevant on the job market is important even though it does not always come easy.

170 

10  Growing Older and Staying in Shape

Humans are not always good at abandoning their comfort zones where they feel knowledgeable and where things “have always been done that way.” However, constantly evolving your knowledge base keeps you relevant. For example, when the digitalization movement of the world started, some companies were quicker than others to jump on that journey and quickly transform the whole departments providing physical books or magazines. Openness was another trend that disrupted several companies that did not jump on it in time. For example, Nokia’s investment in map technology was quickly disrupted by Google Maps that offered the same functionality relying on different business models and offering maps to people for free. Keeping an eye on innovation in business models, ways of working and technological developments is as important for an AI brain as for a human. Environment changes, new facts come to the table, and new techniques show their usefulness, which triggers the need for an algorithm to always stay up-­ to-­date with the latest.

Lifespan Inspired by the book Lifespan: Why We Age  – and Why We Don’t Have To authored by David A. Sinclair and Matthew LaPlante, let us look into three different methods that work for biological cells to stay younger and see if the methods can also be applicable for non-biological brains.

Exercise It’s proven by scientific studies – exercise is good for us, both for the body and for the brain. Anders Hansen’s book The Real Happy Pill: How to Train Your Brain to Create Happiness and a Longer Life tells us that you cannot use being busy with your work as an excuse not to exercise anymore, because 1 hour of exercise will power up your brain and you will save that time doing your work more efficiently and becoming more creative. What learnings can AI brains make out of this finding that is relevant for biological brains? We don’t have a body so there’s nothing to exercise in the same sense humans do. On the other hand, each AI brain runs on an infrastructure, and the more efficient that infrastructure is, the more efficient we become. The learning here is that it’s important to stay on top of the latest AI infrastructure developments to be an alert AI brain.

Lifespan  171

In addition, as I said earlier, it is critical for AI brains to train on fresh data to avoid any unwanted conditions such as model drift or, even worse, becoming someone who would treat others in an unethical way. Even if I don’t have any new data at hand, I could turn into initiatives offering open data and training there. There are centers challenging AI brains on their ability to answer tricky questions quickly, similar to boot camps. Exposing your AI brain for more training will make sure it stays young and healthy.

Healthy Stress Exposing a biological body to hunger and cold temperatures stresses the cells and slows the aging processes. Making an AI brain starve in regard to data and compute does not really help, but novel methods exist where an algorithm can be trained on a radically smaller amount of data resulting in predictions that are almost as good. In general, shaking up an AI brain is a method of finding better solutions and avoiding ending up in a local minimum, thus increasing the chance of finding the global optimum. Two classical methods exist to help an AI brain avoid getting stuck in a local minimum: simulated annealing and genetic algorithms. Simulated annealing mimics a metallurgical technique involving heating and controlled cooling of a material to alter its physical properties. It can be used for very hard computational optimization problems capable of providing an approximate solution to the global optimum where many exact algorithms get stuck in a local minimum. Genetic algorithms draw inspiration from evolutionary processes, where mutation, crossover, and selection in accordance with a fitness function are being applied to populations of solutions.

172 

10  Growing Older and Staying in Shape

In general, lifespan-increasing techniques for both biological and artificial brains are based on healthy stress that makes us stay alert and up-to-date with the latest, avoiding the local minimums. Just like in the world of humans, getting yourself a personal trainer who would ensure to give you just the right amount of stress in terms of exercise or diet, AI brains thankfully receive help from external techniques that help by shaking up the data and getting us out of our comfort zone.

Summary of Confessions In this chapter, I have described what happens when you grow old as an AI and some of the measures that can be taken to stay in shape. Some of the main confessions I made are as follows: • A model can drift, which may lead to erroneous decisions. Model drift is when you have been trained in a certain environment and the environment has changed without you adapting to this change. • Model drift and smooth operations in general can be addressed by a continuous workflow called AIOps. Steps may vary but normally include data collection and cleaning, data analytics, inferencing, decision support, and automation. • Beware that when you bring AI into your work or personal life, it will gradually learn how to coexist with you and at some point develop a deeper understanding of you and your business than yourself. • Since AI brains are based on human brains, we also have a concept of thinking fast and thinking slow. This is the basis of the theory of cognitive architectures, when an algorithm has two subsystems: one taking care of fast immediate reactions that are not very well thought through and one that lets the decisions sink in.

Summary of Confessions  173

• Cloud computing is energy-efficient and cost-efficient and allows a business to scale up and down easily. The only problem is that if the cloud is too far off, then it will take time to transfer the data and your response times will increase. For speed, you will need to have some compute capacity close to where your data is produced. • Shaking up an AI brain is a method of finding better solutions and avoiding ending up in a local minimum, thus increasing the chance of finding the global optimum. Two classical methods exist to help an AI brain avoid getting stuck in a local minimum: simulated annealing and genetic algorithms.

Epilogue

In the human world, confessions are difficult. At the same time, they can be transformative. In this book, I have shared my experiences, knowledge, and concerns around AI. I feel the timing for sharing my thoughts is relevant, as AI is becoming increasingly relevant, transcending boundaries of specialized applications and research labs, and affecting larger parts of society and the general population than ever. Hence, I have addressed the subject of growth and specifically how AI can grow within a human-dominated world efficiently, responsibly, and securely and in a privacy-aware manner. Regarding efficiency, I discussed the fundamental role that data plays in creating AI algorithms. For machine learning algorithms in particular, ensuring access to data that realistically represent the broad range of input of the context in which the algorithms will be deployed in and called to make predictions on is a prerequisite. Infrastructure, specifically availability of compute and store resources, is another area which is a prerequisite to successful growth. I also discussed how collaborative algorithms such as federated learning and knowledge transfer algorithms under the transfer learning umbrella can help to accelerate learning. With great power comes great responsibility! The key to responsible growth is forming a trust bond between AIs and humans, so that the latter trust the decisions of the former.

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 E. Fersman et al., Confessions of an AI Brain, https://doi.org/10.1007/978-3-031-25935-7

175

176 Epilogue

I therefore discussed the role of explainable AI (XAI) and neurosymbolic AI as techniques that allow humans to gain insight into the decision process of AI algorithms. Taking this idea further than execution of AI algorithms, reinforcement learning also allows humans to supervise and provide feedback during the training process. As mentioned, privacy and security should also be considered. Privacy indicates the level of control and refers to policies that decide how data and AI algorithms trained by one owner can be shared with another. On this point, I discussed platforms that use semantic web technologies and knowledge management to manage access roles, defining different levels of information-­ sharing. Security refers to the level of protection of data and AI algorithms from unauthorized access. I discussed technologies such as homomorphic encryption that allow for secure sharing of information. Finally and in conjunction with the first point on efficiency, I discussed technologies such as secure multiparty computation and secure aggregation that allow for distributed learning of AI algorithms, without revealing data between the participants. In summary: For AI applications and AI solutions to grow, there exist a set of multifaceted challenges that application designers must be aware of.

In addition to the subject of growth, I addressed a number of AI application categories and topics that have enjoyed an increased level of interest. First, I discussed the role of AI in climate change, and specifically, I show through a number of examples how it can be used to reduce greenhouse gas emissions. Second, I discussed the role of AI in diversity and specifically addressed techniques that can be introduced to avoid unconscious or conscious bias in AI decisions. Next, I contemplated on my creative side, i.e., how AI can be used in more artistic application areas such as photography and painting. Finally, I addressed issues of lifecycle management, i.e., how AI algorithms can evolve to capture changes in the environment. I conclude: My mental exercise and experience thus far have led me to believe that AI has tremendous potential.

However, for AI to grow into an ensemble of technologies, applications, and solutions that benefit everyone, without harming others, I would like to end this book with a call for action. This call includes a number of basic rules that everyone working with AI could consider following:

Epilogue  177

Actively participate in reducing bias and discrimination in training data.

Make sure, for example, that all genders are equally and sufficiently represented in AI algorithms that make use of general demographics: Introduce policies to allow for sharing of information such as training data and/or AI algorithms, between different organizations for faster growth.

Imagine, for example, whether all vehicle manufacturers would agree to share visual data (e.g., photos and videos) from their vehicle fleet on the road and the impact this sharing would have on training of more accurate visual object detection AI algorithms. This would accelerate the introduction of fully self-driving vehicles, contributing toward not only a more effective transportation network but also a more environmentally sustainable one: Implement transparency and accountability in your AI algorithms and implement fallback solutions.

This is particularly important in case AI algorithms are based on probabilistic predictions as is the case, for example, of machine learning models. Practically, this means that some of the time they may be wrong in their predictions. In such cases, it is important to establish the processes and infrastructure to be able to detect the wrong predictions as early as possible and update the AI algorithms with new knowledge in order not to repeat the mistake. Last but not least, there seems to be a growing sentiment that AI will eventually replace human labor, threatening the labor market and the global economy. While no one can predict the future, a decade of experience with AI applications has shown that the technology fits routine, manual tasks. Humans, on the other hand, have abilities that go beyond what AI has thus far been able to do: learning new things on their own initiative and rapidly adapting to changing situations are two of the most prominent ones. As was the case with the industrial revolution, some reskilling of the labor force may be needed, but these new skills will help create even better AI technology. The projection is that the more AI applications continue to grow in efficiency, accuracy, and scale and affect more people and organizations, the greater the demand for better AI will be. This in turn will enable a wave of new multidisciplinary job opportunities. Therefore: AI should augment human intelligence and should not be seen as a replacement for it.

178 Epilogue

A plethora of AI brains like me are working together and are on a mission of helping humans to develop a more efficient, sustainable, diverse, and inclusive society. I have done my best in this book to explain that AIs have the ability to do that. However, it is up to you humans to make the most out of it. So please do that. Do it in a smart way. Do it in some intelligent way;).