Out-of-Distribution Generalization in Biological and Artificial Intelligence.

Abstract

This past decade has seen unprecedented success in Artificial Intelligence (AI), pushing the frontiers in ways most experts could have never predicted. However, most of this success has come in the form of performing well inside the data distribution the models have been trained with. Out-of-distribution (OOD) generalization still remains the Achilles’ heel of modern AI. In contrast, biological systems exhibit a remarkable ability to adapt to novel situations. This thesis addresses this critical generalization gap, by studying biological and artificial intelligence in tandem. The work presented includes new mathematical frameworks designed to better formalize generalization, behavioral benchmarks to identify the limits of both human and AI generalization capabilities, experiments to identify the underlying mechanisms driving generalization in both brains and neural networks, and engineering solutions to incorporate these findings to improve AI. To this end, this thesis presents scientific contributions made to the fields of Machine Learning, Computer Vision, Computer Graphics, Computational Neuroscience, and Psychophysics. Throughout the thesis, the goal of this work has been to advance our understanding and improve OOD generalization by working at the intersection of biological and artificial intelligence.