Unsupervised Program Synthesis: Hierarchy and Perception (via Zoom)

Abstract: How can we discover interpretable patterns and regularities in datasets? I will discuss a specific approach to this problem, called *unsupervised program synthesis*, which seeks to construct generative programs that reconstruct the input data. This talk will give two applications of this framework:

1. Discovering language patterns in linguistics problems. The talk will show that by doing hierarchical Bayesian inference across many linguistics problems for many different languages, the unsupervised learner can better discover regularities in how natural languages build words.

2. Applying unsupervised program synthesis to raw perceptual data. The end of the talk will show early work on applying the framework to learning game-rules from pixel input. This second part is joint work with Richard Evans.

Bio: Kevin Ellis is an assistant professor in Computer Ccience at Cornell, working in artificial intelligence and programming languages. Previously he was a research scientist at Common Sense Machines and a graduate student in cognitive science at MIT.