Similarities and Differences: Comparing Humans and Deep Networks
Thursday 19 May 2022
Tomas Griffiths, Henry R. Luce Professor of Information Technology, Consciousness, and Culture, Departments of Psychology and Computer Science, Princeton University, USA
Deep networks perform remarkably well in predicting human judgments, but also display systematic differences in the representations they form. I will summarize some work we have done in my lab on using deep networks to predict human similarity judgments and describe a strategy we have developed for identifying cases where a network mimics human behavior but actually has a very different underlying representation. This approach, based on constructing neural network “metamers,” can also be used to identify interventions that bring the inductive biases of machines into closer alignment with those of people.
Short Bio: Tom Griffiths is a Cognitive Scientist interested in developing mathematical models of higher-level cognition; and understanding the formal principles that underlie our ability to solve the computational problems we face in everyday life. His research interests cover a wide range of topics, from perception to decision making and reasoning and cultural evolution, often comparing human behaviour to idealized models to make scientific inferences. He received his PhD in Psychology from the University of Stanford. Since then, he has held permanent positions at both Brown and Berkley and since 2018 at Princeton where he is the Henry R. Luce Professor of Information Technology, Consciousness and Culture.