Active Learning for Neurosymbolic Program Synthesis
summary
Proposes a concrete active-learning pipeline for neurosymbolic synthesis that models neural uncertainty with constrained conformal evaluation and avoids exponential enumeration by a lightweight bidirectional abstract interpretation scheme. Implements the method in SmartLabel and evaluates on 112 tasks across image editing, image search, and visual-arithmetic domains (≈200 inputs, ≈800 questions, ≈1,500 candidate programs each). Reports 98% recovery of the ground-truth program with under five interaction rounds on average, versus 65% and 58% for two prior baselines.
AI-written summary · about
chapters (12)
- 0:00Neuro-symbolic program synthesis intro
- 1:47Active learning challenges
- 3:10Conformal prediction for uncertainty
- 4:54Neuro-symbolic active learning loop
- 6:18Visual arithmetic example
- 7:41Distinguishing candidate programs
- 8:23Question selection and pruning
- 9:50Efficient constraint evaluation
- 12:16SmartLabel tool evaluation
- 13:41Ablation study and performance
- 14:02Conclusion and future work
- 14:24Q&A
related talks
01h 10mNeurosymbolic Program Synthesis: Bridging Perception and Reasoning in Real-World(…)
01h 00mBetter Learning through Programming Languages
38mProgram Synthesis: A Dream Realized?
39mActive learning in the interactive python environment
42mReasonable AI
15mAbstraction Refinement-guided Program Synthesis for Robot Learning from Demonstrations
51mEfficient Deep Learning with Humans in the Loop
30mNeural Architecture Search using Property Guided Synthesis