Scherer / Candrian / Ag You & AI: A Guide to Understanding How Artificial Intelligence Is Shaping Our Lives

1. Auflage 2023
ISBN: 978-3-7528-1361-6
Verlag: BoD - Books on Demand
Format: EPUB
Kopierschutz: 6 - ePub Watermark

E-Book, Englisch, 277 Seiten

As we increasingly integrate artificial intelligence (AI) into our everyday lives, many pressing questions remain: What exactly is AI, and how does it differ from human intelligence? How will AI influence our future, and what challenges must we overcome to develop ethical AI? Explore the exciting world of AI and its impact on our daily lives and society with this ultimate guide. Dr. Anne Scherer and Dr. Cindy Candrian reveal everything about the latest scientific findings on the big questions of AI. Discover the evolution of AI and how unconscious perceptions can influence our trust in it. Learn more about the creativity of machines and how our data is used by AI. With this book, you will learn how to harness the power of AI to make better decisions and what to pay particular attention to, so you don't inadvertently get manipulated, deprived of your abilities, or led to discriminatory decisions. Are you ready to unlock the secrets of "You & AI"? Then this book is perfect for you.

Anne Scherer is a true pioneer in the fields of consumer psychology and technology. For over a decade, she has been on a mission to discover how new technologies are transforming the way we interact with businesses. As an Assistant Professor of Quantitative Marketing at the University of Zurich and co-founder of Delta Labs AG, Anne delves deep into the fascinating world of AI, robo-advisors, and conversational interfaces. Driven by her passion for "better tech," Anne supports startups, companies, and NGOs in shaping our AI-driven future. She has been a member of the World Economic Forum's Global Future Councils and is known for her groundbreaking research, published in top academic journals and major media outlets. Her TEDxTalk, in which she discusses why we are more honest with machines, has already been viewed over 1.8 million times. Before joining the University of Zurich, Anne conducted research at ETH Zurich and earned her PhD with honors from the Technical University of Munich. Anne is not only a scientist but also a true adventurer. She has traveled the world, from diving in the Red Sea to cycling on Easter Island, and these experiences have enriched her perspective. In this way, Anne contributes to shaping a technology-driven future that is as exciting as her own adventures. Anne lives with her partner in Zurich.

Autoren/Hrsg.

Scherer, Anne

Candrian, Cindy

Ag, Delta Labs

Leseproben

Buzzword Bingo Explained: From Machine Learning to Generative AI Obviously, the basis of intelligence is—you guessed it—learning. Learning means we improve our performance in the future after we have observed and taken in some information about the world around us. This is no different for today’s AI. In the world of AI, learning often means that we have a vast collection of input-output pairs, from which an underlying function is derived so the model can predict the output for new incoming input. Sound complicated? Let’s illustrate that with an easy example. Say we have a supercool collection of cute cat and dog pictures that show all different kinds of cats or dogs in them in all kinds of different situations. This is our input. As an output, we tell the model that these are either cats or dogs in the pictures. Now from this collection, we want to model to learn to recognize and distinguish cats and dogs. But not just those in our collection! This is no game of memory. That would be too easy. No; we want to feed the model a completely new cat image, and from this new input we want it to be able to predict that this is indeed a cat—and not our neighbor’s tiny chihuahua. You may wonder, why go through all this hassle? Collecting all these pictures and tagging all the cats and dogs in there so the program can learn. Why not directly tell the machine all the steps to take to come to solve a problem or come to a conclusion? Just like math lessons, couldn’t we just simply let the program know all our math rules and functions so there is no more need for learning? Right . . . and wrong. This may work for simple tasks and problems. A lot of early AI research tried to teach the machine this way. And while they did have some success on smaller problems, they all failed miserably when it came to applying the models in the real world. And our world is messy, complex, and full of unknowns. So to be able to use AI for bigger and more complex problems, we need it to learn. Think about it. We simply cannot anticipate all situations the program will be in. Consider a self-driving car. Now think about all cars on the road worldwide; the weather and road conditions; the cars’ wear and tear; and other cars, people, or obstacles on the road. Quickly we must realize that it is simply impossible to prepare an AI system for all possible situations. So it needs to learn. Also, there may be changes with time. Consider a program designed to predict the stock market. Now a global pandemic comes along. It needs to be able to adapt when conditions change from boom to gloom to make good predictions. So it needs to learn. And most of all, sometimes we have no clue ourselves! Just consider our cat pictures. While it may be easy for us to say that these are all cats, even the best programmers might find it hard to boil that down into an algorithm—unless it is self-learning. So in short, for AI to become truly intelligent, it needs to be able to learn! This brings us to the next point. How does AI learn? You may have come across many of the terms here: machine learning, supervised learning, unsupervised learning, reinforcement learning, deep learning. Basically these are all different ways of a machine, learning. Machine learning is the broadest concept of them all. It basically refers to the idea that our algorithm is able to learn. Remember how we said that for very simple problems, we can tell a machine directly what to do. So it would look something like this: “if this” . . . “then do this.” Clearly there is no learning from the computer’s side. It is us telling it exactly what we know. Now consider a more complex task, such as teaching our machine to tell the difference between cats and dogs. As you can imagine, it would be extremely hard and complex to narrow this down to an if-then sequence. So we want the machine to learn. Or put differently, we use machine learning. Although it may sound complicated, the term basically refers to a set of approaches that all share the idea that a machine will learn from the data it is provided to improve its performance on a given task. That’s it! That wasn’t hard after all, right? And you have all seen them before, too! Familiar examples are the Netflix recommendation system, Snapchat filters, Google Maps, and Spotify-generated playlists. All use AI models based on machine learning! Supervised, unsupervised, and reinforcement learning are all forms of machine learning. The only thing that differs among these types of learning is the way feedback is provided to the system. Very intuitively, supervised learning means we give our algorithm all the feedback we can give. In short, learning is supervised. What does this look like? Well, we simply use so-called “labeled” data. This labeled data contains both input and output, meaning the output is already known. Think about our cat and dog images. Instead of just feeding cat and dog images into the program, we also tell it if there is a cat or a dog in each picture by labeling the images as either “cat” or “dog.” So supervised learning is like having a teacher who shows you examples and tells you the correct answers. The more examples you see, the better you get at solving problems on your own. Once you get the logic, you receive new examples and solve them based on what your teacher taught you. Clearly this labeling is time- and labor-intensive. So researchers tried to find ways around having experts painstakingly annotate data for supervised learning—and they have been successful! In 2022, a group of researchers were able to train an AI model called CheXzero to spot diseases on chest X-rays with medical reports that experts had written in natural language. And while you may think this hurts performance, it did not! In fact, the self-supervised model outperformed the supervised models with fully labeled data. This paves the way for self-supervised AI models that no longer need any data with explicit annotations and makes machine learning even faster! Let’s move on to unsupervised learning, which is, as the name implies, learning without explicit feedback. No supervision. This means that the program is not given any labeled data. Instead, the machine is left on its own to group the unsorted information by finding similarities, differences, and patterns in the data. It’s like a detective trying to solve a mystery. It doesn’t have all the clues up front, so it has to gather evidence and try to piece together what’s going on. Consider our cat and dog images again. This time we do not tell which image depicts a dog and which a cat. The idea of unsupervised learning is that by carefully examining each image, the machine can identify the clues that separate cats and dogs, such as the length of their tails, the presence of retractable claws, or the number of whiskers. After analyzing the evidence, the machine can group the images into categories, but it can’t quite tell us if they are cats or dogs. Obviously this approach does not only apply to images of cats and dogs. In fact, this approach is often used in marketing, where marketers need to identify different customer segments. Through unsupervised learning, the AI can cluster customers together that share important features and are sufficiently dissimilar to other groups of customers. Then there is reinforcement learning. You may have heard of this before. And you are right, reinforcement is a very common way we humans learn, too. When you think about reinforcement, think about rewards and punishments. We often use those in our day-to-day lives to reinforce positive behaviors in others while weakening negative ones. So we praise kids for eating their veggies while withholding the dessert if they don’t. How does this relate to AI? Well, we can provide feedback to our machine in the exact same manner so it may learn the best behavior or actions. For example, we can reward it with two points if it wins a chess game and deduct two if it loses. Similarly, we can add points if it correctly identifies our cats and dogs in the images. The goal for the program then is to maximize the rewards and minimize the punishments. The machine achieves this by deciding which action prior to the reinforcement (positive and negative) was most responsible for it. It will then show those actions more often that led to a reward and reduce those that led to a punishment. So if it was a kid at the dinner table, it would eat those veggies to gather more praise and eat even more to make sure it no longer misses out on that yummy dessert! As you can imagine, the more data we feed our machine, the better for learning. So it is no coincidence that the rise of big data and advancements in computing power have led to even better ways of learning. This is where another form of machine learning, called deep learning, comes in. Deep learning is a type of artificial intelligence that’s modeled after the way our brains work. Why is it called deep? Well, it’s the...