Explore a comprehensive lecture on DreamCoder, a system for bootstrapping inductive program synthesis through wake-sleep library learning. Delve into the premise of program induction, its applications in visual programming and code writing, and the concept of library learning as Bayesian inference. Examine how neural recognition models guide the search process and discover the domains where DreamCoder excels, including LOGO Turtle Graphics. Investigate the system's dreaming process before and after learning, its application in planning tower construction, and the synergy between recognition models and library learning. Gain insights into the growth of languages for vector algebra, physics, and recursive programming, concluding with valuable lessons from this innovative approach to program synthesis.
Dreamcoder- Bootstrapping Inductive Program Synthesis With Wake-Sleep Library Learning
Overview
Syllabus
Intro
The premise of program induction
Why program induction?
Visual programs
Learning to write code
Library learning as Bayesian inference
Library learning as neurally-guided Bayesian inference
Abstraction Sleep: Growing the library via refactoring
Neural recognition model guides search
DreamCoder Domains
LOGO Turtle Graphics - learning an interpretable library
What does DreamCoder dream of7 (before learning)
What does DreamCoder dream of7 (after learning)
What does DreamCoder dream of (after learning)
Planning to build towers
Dreams after learning
Learning dynamics
Synergy between recognition model and library learning
Evidence for dreaming bootstrapping better libraries
Growing languages for vector algebra and physics
Growing a language for recursive programming
Lessons
Taught by
Simons Institute
