Learning to Learn Concepts as Programs: Hierarchical Bayes and Amortised Inference
Abstract
The PI’s investigation looks to impute patterns in perceptual stimuli, and learn new concepts remarkably efficiently in terms of both how much they can learn from minimal data, and the speed with which they achieve this. Despite impressive recent advances in neural networks for machine learning, these models typically require thousands of examples to learn about a new domain, and moreover lack the rich, interpretable representational structure that characterizes even the simplest human concepts. Meanwhile symbolic models of concept learning, while interpretable, have suffered from brittleness when applied to new domains, and are often severely limited by the amount of computation they require to learn simple concepts. In the proposed work, the PI will unite these two classes of model within the framework of Bayesian Program Learning (BPL), aiming to combine the strengths of both.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 14, 2022
- Source ID
- FA95501910269
Entities
People
- Joshua B. Tenenbaum
Organizations
- Air Force Office of Scientific Research
- Massachusetts Institute of Technology
- United States Air Force