Probabilistic Neural Programs
This work addresses the problem of flexible program specification and inference in AI, offering a domain-specific incremental improvement for tasks like diagram question answering.
The paper tackles program induction by introducing probabilistic neural programs, a framework that integrates neural networks with probabilistic choice operators, and achieves nearly double the correct program execution rate compared to a baseline on a diagram question answering task.
We present probabilistic neural programs, a framework for program induction that permits flexible specification of both a computational model and inference algorithm while simultaneously enabling the use of deep neural networks. Probabilistic neural programs combine a computation graph for specifying a neural network with an operator for weighted nondeterministic choice. Thus, a program describes both a collection of decisions as well as the neural network architecture used to make each one. We evaluate our approach on a challenging diagram question answering task where probabilistic neural programs correctly execute nearly twice as many programs as a baseline model.