The Google DeepMind group did some work on Atari video games. They taught, they used reinforcement learning just like in AlphaGo to teach the machine how to play Breakout. But then another group did an experiment where they took the paddle and they moved it up to pixels. Now the program could not play the game at all because it hadn't abstracted the notion of a paddle as an object. As if we would see the world and not see objects. We may have to build some things into our AI programs.
Artificial intelligence is better than humans at playing chess or go, but still has trouble holding a conversation or driving a car. A simple way to think about the discrepancy is through the lens of “common sense” — there are features of the world, from the fact that tables are solid to the prediction that a tree won’t walk across the street, that humans take for granted but that machines have difficulty learning. Melanie Mitchell is a computer scientist and complexity researcher who has written a new book about the prospects of modern AI. We talk about deep learning and other AI strategies, why they currently fall short at equipping computers with a functional “folk physics” understanding of the world, and how we might move forward.
Support Mindscape on Patreon.
Melanie Mitchell received her Ph.D. in computer science from the University of Michigan. She is currently a professor of computer science at Portland State University and an external professor at the Santa Fe Institute. Her research focuses on genetic algorithms, cellular automata, and analogical reasoning. She is the author of An Introduction to Genetic Algorithms, Complexity: A Guided Tour, and most recently Artificial Intelligence: A Guide for Thinking Humans. She originated the Santa Fe Institute’s Complexity Explorer project, on online learning resource for complex systems.
See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.