ENS, Room Dussane, 45 rue d'Ulm
In 1986, Dave Rumelhart, Geoff Hinton, and I began the first chapter of Parallel Distributed Processing, a two-volume work proposing neural network based models of human cognition, with the question ‘Why are people smarter than machines?’ At the time, people were far better than existing machine systems in many ways. Since then, machines have come a long way, and many of their successes rely on the kinds of mechanisms we promoted in the PDP volumes a third of a century ago. Neural-network based artificial systems now dominate humans at games like Chess and Go, and they have achieved breakthroughs in vision, language, any many other domains. Yet, it seems clear that these systems have not yet captured important aspects of human intelligence. I will compare human and artificial neural networks and point out some of the ways in which human still exceed our current machine approaches, focusing on a current project that illustrates some of the key differences between human and contemporary machine intelligence. Do these shortcomings and differences mean we need a radically different approach? In the last part of the talk, I will share my thoughts on this question, and give suggestions for next steps toward addressing the limitations of our current machine systems.
Nam, A. J. & McClelland, J. L. (2021). What underlies rapid learning and systematic generalization in humans. Draft dated June 23, 2021. [PDF]
McClelland, J. L. (in press). Could the AI of our dreams ever become reality? In Vernallis, C., Rogers, H., Leal, J., and Kara, S. (Eds.), Cybermedia: Science, Sound and Vision. New York, NY: Bloomsbury Academic. [PDF]
McClelland, J. L., Hill, F., Rudolph, M., Baldridge, J., & Schuetze, H. (2020). Placing language in and integrated understanding system: Next steps toward human-level performance in neural language models. Proceedings of the National Academy of Sciences, 117(42), 25966-25974. DOI: 10.1073/pnas.1910416117. [ PDF]