DreamCoder: Growing generalizable, interpretable knowledge with wake-sleep Bayesian program learning
Type: ./types/snapshot.md
Author: Kevin Ellis, Catherine Wong, Maxwell Nye, Mathias Sable-Meyer, Luc Cary, Lucas Morales, Luke Hewitt, Armando Solar-Lezama, Joshua B. Tenenbaum Source: https://arxiv.org/abs/2006.08381 Date: June 15, 2020
ArXiv ID: 2006.08381
Subject Categories: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
Abstract
The paper presents DreamCoder, a system that learns to solve problems through program writing. The system builds expertise by developing programming languages that express domain-specific concepts alongside neural networks that direct the search for programs within those languages.
The approach employs a "wake-sleep" learning algorithm that iteratively extends the language with new symbolic abstractions and trains neural networks on both imagined and replayed problems. The system tackles inductive programming tasks and creative endeavors like picture drawing and scene construction. Notably, it independently rediscovers foundational concepts including "functional programming basics, vector algebra and classical physics, including Newton's and Coulomb's laws."
The resulting knowledge representations are compositional, building hierarchically from previously learned concepts. This approach yields interpretable and transferable multi-layered symbolic structures that scale flexibly with accumulated experience.