Better Learning through Programming Languages
summary
Argues that combining program synthesis with ML lets search scale and tolerate noise by directing search with a network rather than exhaustive enumeration, demonstrating a neural-guided search that feeds interpreter-produced program states (state-as-input) to the model, applied to geometric and 3D shape reconstruction, number-word translation and DSL learning (NeurIPS 2020, ICLR 2020). Shows a teacher-student loop for hybrid controllers that extrapolates better than pure RL; prototypes work but miss small details.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:53Program synthesis overview
- 3:37Broadening applications
- 8:00Scientific and design use
- 12:56Current challenges
- 15:55Neural guided search
- 20:14Bottom-up synthesis
- 23:57Combining PL and neural
- 29:57Symbolic generalization
- 36:49Teacher-student learning
- 47:23Scientific discovery
- 49:32Q&A session
related talks
01h 10mModels and Programs: Better Together
18mActive Learning for Neurosymbolic Program Synthesis
01h 10mNeurosymbolic Program Synthesis: Bridging Perception and Reasoning in Real-World(…)
26mHow Machine Learning Can Help
22mA Hybrid Meta-Learning Framework for Adaptive Safe Controller Synthesis of Dynamical(…)
28mPreference Learning from Minimal Human Feedback for Interactive
42mReasonable AI
54mProgramming Language Design and Implementation for the Machine Learning Era: A Personal(…)