<div dir="ltr"><p style="box-sizing:inherit;margin:0px;padding:0px;outline:0px;border:0px;background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial;vertical-align:baseline;font-size:18px;color:rgb(50,50,50);font-family:Georgia,serif">I remember that <em style="background:transparent;box-sizing:inherit;margin:0px;padding:0px;outline:0px;border:0px;vertical-align:baseline">Gödel, Escher, Bach</em> was Eliezer Yudkowsky's favorite book. I hope that his think tank is doing well. </p><p style="box-sizing:inherit;margin:0px;padding:0px;outline:0px;border:0px;background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial;vertical-align:baseline;font-size:18px;color:rgb(50,50,50);font-family:Georgia,serif"><br></p><p style="box-sizing:inherit;margin:0px;padding:0px;outline:0px;border:0px;background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial;vertical-align:baseline;font-size:18px;color:rgb(50,50,50);font-family:Georgia,serif">"The Pulitzer Prize-winning book <em style="box-sizing:inherit;margin:0px;padding:0px;outline:0px;border:0px;background:transparent;vertical-align:baseline">Gödel, Escher, Bach</em> inspired legions of computer scientists in 1979, but few were as inspired as <a href="https://www.santafe.edu/people/profile/melanie-mitchell" style="box-sizing:inherit;color:rgb(101,101,101);text-decoration-line:none;border-bottom:1px dotted rgb(101,101,101)" target="_blank">Melanie Mitchell</a>. After reading the 777-page tome, Mitchell, a high school math teacher in New York, decided she “needed to be” in artificial intelligence. She soon tracked down the book’s author, AI researcher Douglas Hofstadter, and talked him into giving her an internship. She had only taken a handful of computer science courses at the time, but he seemed impressed with her chutzpah and unconcerned about her academic credentials.</p><p style="box-sizing:inherit;margin:30px 0px 0px;padding:0px;outline:0px;border:0px;background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial;vertical-align:baseline;font-size:18px;color:rgb(50,50,50);font-family:Georgia,serif">Mitchell prepared a “last-minute” graduate school application and joined Hofstadter’s new lab at the University of Michigan in Ann Arbor. The two spent the next six years collaborating closely on <a href="https://www.sciencedirect.com/science/article/abs/pii/0167278990900865" style="box-sizing:inherit;color:rgb(101,101,101);text-decoration-line:none;border-bottom:1px dotted rgb(101,101,101)" target="_blank">Copycat</a>, a computer program which, <a href="https://theputnamprogram.files.wordpress.com/2012/03/dhfa_chp5.pdf" style="box-sizing:inherit;color:rgb(101,101,101);text-decoration-line:none;border-bottom:1px dotted rgb(101,101,101)" target="_blank">in the words of its co-creators</a>, was designed to “discover insightful analogies, and to do so in a psychologically realistic way.”</p><p style="box-sizing:inherit;margin:30px 0px 0px;padding:0px;outline:0px;border:0px;background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial;vertical-align:baseline;font-size:18px;color:rgb(50,50,50);font-family:Georgia,serif">The analogies Copycat <a href="https://arxiv.org/abs/2102.10717" style="box-sizing:inherit;color:rgb(101,101,101);text-decoration-line:none;border-bottom:1px dotted rgb(101,101,101)" target="_blank">came up with</a> were between simple patterns of letters, akin to the analogies on standardized tests. One example: “If the string ‘abc’ changes to the string ‘abd,’ what does the string ‘pqrs’ change to?” Hofstadter and Mitchell believed that understanding the cognitive process of analogy—how human beings make abstract connections between similar ideas, perceptions and experiences—would be crucial to unlocking humanlike artificial intelligence.</p><span style="color:rgb(50,50,50);font-family:Georgia,serif;font-size:18px"></span><p style="box-sizing:inherit;margin:30px 0px 0px;padding:0px;outline:0px;border:0px;background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial;vertical-align:baseline;font-size:18px;color:rgb(50,50,50);font-family:Georgia,serif">Mitchell maintains that analogy can go much deeper than exam-style pattern matching. “It’s understanding the essence of a situation by mapping it to another situation that is already understood,” she said. “If you tell me a story and I say, ‘Oh, the same thing happened to me,’ literally the same thing did not happen to me that happened to you, but I can make a mapping that makes it seem very analogous. It’s something that we humans do all the time without even realizing we’re doing it. We’re swimming in this sea of analogies constantly.”</p><p style="box-sizing:inherit;margin:30px 0px 0px;padding:0px;outline:0px;border:0px;background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial;vertical-align:baseline;font-size:18px;color:rgb(50,50,50);font-family:Georgia,serif">As the Davis professor of complexity at the Santa Fe Institute, Mitchell has broadened her research beyond machine learning. She’s currently leading SFI’s <a href="https://santafe.edu/events/foundations-intelligence-natural-and-artificial-sy" style="box-sizing:inherit;color:rgb(101,101,101);text-decoration-line:none;border-bottom:1px dotted rgb(101,101,101)" target="_blank">Foundations of Intelligence in Natural and Artificial Systems</a> project, which will convene a series of interdisciplinary workshops over the next year examining how biological evolution, collective behavior (like that of social insects such as ants) and a physical body all contribute to intelligence. But the role of analogy looms larger than ever in her work, especially in AI—a field whose major advances over the past decade have been largely driven by deep neural networks, a technology that mimics the layered organization of neurons in mammal brains.</p><p style="box-sizing:inherit;margin:30px 0px 0px;padding:0px;outline:0px;border:0px;background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial;vertical-align:baseline;font-size:18px;color:rgb(50,50,50);font-family:Georgia,serif">“Today’s state-of-the-art neural networks are very good at certain tasks,” she said, “but they’re very bad at taking what they’ve learned in one kind of situation and transferring it to another”—the essence of analogy."</p><div><font size="4"><br></font></div><div><font size="4"><a href="https://www.scientificamerican.com/article/the-computer-scientist-training-ai-to-think-with-analogies/" target="_blank">https://www.scientificamerican.com/article/the-computer-scientist-training-ai-to-think-with-analogies/</a>     </font><br></div><div><font size="4"><br></font></div></div>