We want to build machines that can learn structured generative models: models that learn to generate rich objects like words, pictures, hand-drawn characters, intuitive theories, or grammars. Bayesian Program Learning (BPL) is a framework for learning structured generative models like these. We put together an engineering toolkit for building BPL systems, and show the toolkit working on language learning problems and on graphics programming problems. In the language domain the BPL model solves most of the problems in a standard linguistics textbook. In the graphics programming domain the BPL model learns to convert hand drawings into a subset of LaTeX. The BPL toolkit is a step toward more human-like agents that, like people, flexibly learn generative models.
UPCOMING COG LUNCH TALKS:
11/21/17 - Dian Yu (Rosenholtz Lab)
11/28/17 - Yang Wu (Schulz Lab)
12/05/17 - Melissa Kline, Ph.D.