Statistics over algorithms as a model of human learning: Neyman Seminar
Seminar | September 4 | 4-5 p.m. | 1011 Evans Hall
Steve Piantadosi, UC Berkeley
Human learning differs qualitatively from state of the art machine learning both in scale and power. People are able to discover much richer latent structures in data than are typically captured in statistical models. In particular, people seem able to discover algorithmically sophisticated representations, often real computational processes like computer programs. This ability can be seen in ordinary human learning of areas like mathematics, language, kinship, and logic. I'll present an overview of my research that is aimed at understanding and modeling how human learners solve these kinds of complex and structured learning problems. I'll present models that perform statistical inference over a Turing-complete spaces of computations, and I'll provide empirical evidence that these learning theories accurately capture human patterns of generalization and representation. This approach allows cognitive science and artificial intelligence to address basic questions about what types of knowledge might be "built in" for humans, and how natural and artificial learners could develop the rich systems of knowledge found in adults. More broadly, the ideas used in this work have the potential to inform the discovery of structure, algorithmic processes, scientific laws, and causal relations by using techniques inspired by the remarkable statistical inferences carried out by humans.