DreamCoder: Growing generalizable, interpretable knowledge with wake-sleep Bayesian program learning
What this work does
Alternates solving tasks, compressing found programs into reusable library abstractions, and training a recognition model on replayed and generated tasks.
Connection to discovery
Separates improving a representation language from improving search within that language, making representation growth an explicit learning target.
Evidence
The inspected algorithm uses typed programs and approximate Bayesian library learning; the domain illustrations include list processing, visual programs and physical-law expressions.
Scope and limitations
The initial primitives, task generator, type system and bounded search still constrain discoveries. The consulted sections establish the algorithm, not a replication of its domain results.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.