Neurosymbolic Program Synthesis: Bridging Perception and Reasoning in Real-World(…)
summary
Neurosymbolic Program Synthesis combines pretrained neural primitives (vision models, classifiers, LLMs) with DSLs and symbolic constructs, using LLM-guided top‑down enumerative search plus program analysis (type systems, abstract interpretation, partial evaluation) to prune and repair. Built tools: ImageEye learns image‑search/edit programs from ~4 examples on average, solving 96% of tasks within a 2‑minute limit (median synth 1s); semantic regaxes beat regex and GPT baselines on F1 by adding LLM semantic checks; Prollex synthesizes robot policies for 80% of benchmarks (65% overall success).
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:04Neurosymbolic program synthesis
- 5:21Deep learning meets program synthesis
- 7:26Neurosymbolic program representation
- 11:18Learning algorithms
- 14:09Image search and editing
- 27:59Text data extraction
- 41:57Robotics with neurosymbolic methods
- 50:46Trust in neurosymbolic learning
- 58:29Future directions
- 1:01:22Q&A
related talks
18mActive Learning for Neurosymbolic Program Synthesis
38mProgram Synthesis: A Dream Realized?
01h 00mBetter Learning through Programming Languages
47mBig Ideas in Program Synthesis
15mAbstraction Refinement-guided Program Synthesis for Robot Learning from Demonstrations
42mReasonable AI
28mSymbolic Reasoning in telecom oriented LLMs
01h 10mModels and Programs: Better Together