In 1979, the Pulitzer Prize-winning book Gdel Escher, Bach inspired a lot of computer scientists, but none were as inspired by Melanie Mitchell. Mitchell, a New York high school math teacher, read the entire 777-page book and decided that artificial intelligence was what she wanted to do. Mitchell quickly tracked down Douglas Hofstadter (an AI researcher) and convinced him to give her an internship. Although she had only taken a few computer science courses at the moment, he was impressed by her intelligence and seemed unconcerned about any of her academic credentials.Mitchell submitted a last-minute application to graduate school and was accepted into Hofstadters' new lab at the University of Michigan, Ann Arbor. They worked closely together on Copycat for six years. This computer program, according to its creators, was created to find insightful analogies and do so in a psychologically real way.Copycat's analogies were made between simple patterns of letters. They are similar to analogies found on standard tests. One example: What happens if the string abc changes from the string abd? How does the string pqrs affect the abc? Hofstadter, Mitchell believed understanding the cognitive processof analogyhow humans make abstract connections between similar ideas and perceptions would be key to unlocking humanlike artificial Intelligence.Mitchell believes analogy goes deeper than pattern matching in an exam. She said that it is about understanding the essence of a situation and mapping it to another situation. Tell me a story, and I will tell you that the same thing happened to me. However, I can create a mapping that makes it appear very similar. This is something we all do without realizing it. This sea of analogies is constantly a constant.Mitchell is the Davis professor for complexity at Santa Fe Institute. This has allowed her to expand her research beyond machine-learning. Mitchell is currently the Foundations of Intelligence in Natural & Artificial Systems project leader at SFI. This will host a series of inter-disciplinary workshops that will examine how intelligence can be achieved through collective behavior, biological evolution, and physical bodies. Her work is dominated by analogy, particularly in AIa field where major advancements over the past decade were largely due to deep neural networks technology, which mimics the layered organization neurons in mammal brains.She said that today's state-of the-art neural networks can be very proficient at some tasks but are terrible at transferring what they have learned from one situation to another. This is the essence of analogy.Quanta spoke to Mitchell about AI and analogies, the field's knowledge so far about them, and what it should do next. This interview has been edited and condensed for clarity.What makes analogy-making so crucial to AI?This fundamental thought process will allow AI to get to the places it wants to go. People believe that AI's ability to predict the future, have common sense and retrieve useful memories is key. Analogy is a key component in all of these areas.We want self-driving cars. But, if they are faced with a situation that is just slightly different from what they have been taught, they don't know what to do. How can we humans learn what to do when we aren't familiar with certain situations? We use analogies from our past experience. This is something we will also need the AI systems of the real world to do.You also wrote that analogy is a fundamental area of AI.It is important to cognition that people don't study it as much. It was important to focus on logic and programming in rules for behavior. This is how early AI worked. People have been focusing on learning from many examples and then inducing new things using the statistics of what they have already learned. It was hoped that the statistics would allow them to abstract and generalize, but this has not worked out as well as they had hoped.A deep neural network can recognize new pictures of bridges by looking at millions of them. It can't abstract the idea of bridge to our notion of bridging gender gaps. It turns out that these networks don't know how to abstract. Something is missing. People are still trying to figure it out.Melanie Mitchell, the Davis professor for complexity at Santa Fe Institute has been working on digital minds for many decades. She believes that AI can never be truly intelligent until it can perform something unique to humans: analogies. Credit: Emily Buder/Quanta Magazine; Gabriella Marks for Quanta MagazineThey won't learn to abstract.New approaches like meta-learning allow machines to learn faster. Self-supervised learning is where GPT-3 can learn to fill out a sentence without one of the words. This allows it to generate language very convincingly. Some would argue that systems such as this will eventually learn how to perform this abstract task if they have enough data. However, I don't think so.This limitation is what you have described as the barrier to understanding AI systems under certain conditions. However, they can mimic understanding but become fragile and unreliable beyond those conditions. What makes you believe analogy is the solution to this problem?I believe that understanding the problem of brittleness will be key to solving it. This is what ultimately leads to the brittleness issue: The systems don't understand the data they are dealing with in a human-like way.The word understand is one those words that everyone agrees on but no one really knows what it means. It almost acts as a placeholder for mental phenomena we don't know how to explain. This mechanism of analogy and abstraction is what I believe is the key to understanding. It is the mechanism that allows us to understand. It allowed us to map something we already knew to something new.An analogy can be used to show how organisms are cognitively flexible and not robotic.Yes, I believe so. Analogy isn't something that humans do only. While some animals can be described as robotic, others are capable of taking previous experiences and mapping them to new experiences. Perhaps it is one way to place a spectrum intelligence on different types of living systems. To what extent are you able to make abstract analogies?One theory that explains why people have such intelligence is that they are so social. It is important to understand the goals of others and to predict their actions. You can do this by analogy with yourself. Put yourself in the shoes of another person and map your mind onto theirs. AI professionals often talk about this theory of mind. It is essentially an analogy.This was the first attempt to do this with a computer. Are there other systems?AI's structure mapping work focused on logic-based representations and mappings between situations. Ken Forbus and others used an analogy that Ernest Rutherford made in 1911, comparing the solar system to an atom. They would have created a set [in a formal language called predicate logic] of sentences that described these two situations. The structure of the sentences was what they used to map them. This idea is powerful and I believe it is right. Humans are more interested in relationships than objects when trying to understand similarity.These approaches have not been successful.These systems largely ignored the whole issue of learning. These words, which were very, very loaded with human meaning, such as the Earth revolves about the sun and electron revolves around nucleus, would be used to map structure mapping. However, there was no internal model for what revolves around. It was a symbol. Copycat was able to handle letter strings well, but we didn't have the answer to the question: How can we scale it up to other domains we care about?Deep learning famously scales quite well. Is it any more effective in producing meaningful analogies than deep learning?Deep neural networks are believed to do this magic between their output and input layers. They should be able recognize different breeds of dogs better than humans. People would then create a large data set to train their neural networks and test them on. Then they would publish papers saying that the data set has strange statistical properties that allow it to learn how to solve these problems without being able generalize. Your machine performs horribly on this new data set, but ours is great.If you have to train your abstraction on thousands or even thousands of examples, then you've already lost the battle. This is not abstraction. Machine learning calls this few-shot learning. This means that you only learn from a small number of examples. This is what abstraction is all about.What is the problem? These approaches could be glued together like Lego blocks.The instruction manual that explains how to do this is not available. However, I believe we must all build them together. This is the frontier of research: What's the key insight out of all these things and how can they compliment each other?Many people are interested in the Abstraction and Reasoning Corpus [ARC], a challenging, one-shot learning task that is based on core knowledge that all humans were born with. We are familiar with the concept of the world being divided into objects and the geometry of space such as something being under or over something else. ARC is a grid of colors that can be transformed into another grid of color. This allows humans to understand the core knowledge of the world.It is an analogy challenge that I view. You are trying to figure out how to describe the change in an image from one to another. There is no way to learn statistical correlations or strange patterns because you only have two examples. This is the hardest part of any system I have mentioned. None of them can handle the ARC data set. It's not a holy grail.Does this mean that an AI will need a body similar to ours if babies are born with this knowledge?This is the million-dollar question. This is a controversial topic that the AI community doesn't agree on. I believe that without some form of embodiment, we won't be able achieve a humanlike analogy in AI. Because some visual problems are difficult to see in three dimensions, having a body is essential. This is because I have lived in the world, moved around and understood spatial relationships. It is not clear if machines have to go through this stage. It will, I believe.Reprinted with permission of Quanta Magazine. This independent publication is part of the Simons Foundation and focuses on research trends and developments in mathematics, the physical, and life sciences.
