Accelerating Human Visual Concept Learning and Boosting Performance via Computational Models of Perception and Cognitio

Visual categorization is ubiquitous in many professions, yet training programs are typically time- and effort-intensive. This work focuses on developing methods to improve human learning and performance on challenging visual categorization tasks, e.g., bird species identification, diagnostic dermatology. As part of the general approach, we infer state-of-the-art psychological embeddings, formal models of the internal representations that individuals use to reason about a domain. Using predictive cognitive models that operate on an embedding, we evaluate different techniques for making training more efficient as well as amplifying an individual's capabilities regardless of experience. In particular, this work concentrates on the value of allowing learners to request clues, the ability to predict exemplar difficulty, and manipulating the order of training trials in order to maximize learning outcomes. Results from a category learning experiment reveal that allowing learners to request clues enables early success at no cost to later performance. Model and behavioral analysis indicate that the difficulty of an exemplar can be accurately predicted without relying on human training data or experts. Results of a category learning experiment suggest that learning outcomes can be improved by arranging the order of trials using a non-conventional scheduling policy. Collectively, the results of this work bring us closer to a world where visual category learning is no harder than it absolutely has to be.