From Perception to Programs: Regularize, Overparameterize, and Amortize
This work addresses the challenge of integrating inductive reasoning with perception for AI systems, though it appears incremental as it builds on existing neurosymbolic methods.
The paper tackles the problem of neurosymbolic program synthesis by developing techniques to combine neural perception with symbolic reasoning, resulting in improved stability for gradient-guided program search.
Toward combining inductive reasoning with perception abilities, we develop techniques for neurosymbolic program synthesis where perceptual input is first parsed by neural nets into a low-dimensional interpretable representation, which is then processed by a synthesized program. We explore several techniques for relaxing the problem and jointly learning all modules end-to-end with gradient descent: multitask learning; amortized inference; overparameterization; and a differentiable strategy for penalizing lengthy programs. Collectedly this toolbox improves the stability of gradient-guided program search, and suggests ways of learning both how to perceive input as discrete abstractions, and how to symbolically process those abstractions as programs.