Learning deep taxonomic priors for concept learning from few positive examples

Human concept learning is surprisingly robust, allowing for precise generalizations given only a few positive examples. Bayesian formulations that account for this behavior require elaborate, pre-specified priors, leaving much of the learning process unexplained. More recent models of concept learning bootstrap from deep representations, but the deep neural networks are themselves trained using millions of positive and negative examples. In machine learning, recent progress in metalearning has provided large-scale learning algorithms that can learn new concepts from a few examples, but these approaches still assume access to implicit negative evidence. In this paper, we formulate a training paradigm that allows a meta-learning algorithm to solve the problem of concept learning from few positive examples. The algorithm discovers a taxonomic prior useful for learning novel concepts even from held-out supercategories and mimics human generalization behavior—the first to do so without hand-specified domain knowledge or negative examples of a novel concept.