Modeling Expertise with Neurally-Guided Bayesian Program Induction

Studies of human expertise suggest that experts and novices “see“ problems differently. Experts not only acquire a body of domain-specific strategies and knowledge, but also learn to quickly identify when those concepts apply to a given problem within the domain. We propose modeling these elements as an iterative process of domain-specific language (DSL) learning, while jointly training a neural network to recognize when learned concepts apply to new problems in the domain. We show that the neural network allows the algorithm to solve problems more accurately and quickly than a model that simply acquires new concepts without learning when to use them. We also examine the implicit, vector-based problem representations learned by the neural network recognition model. Early in training, these representations cluster problems based on surface features, but they increasingly come to reflect abstract relationships between problems as the model acquires domain expertise.