A Machine Learning Approach to Human Visual Learning

University Of Rochester, Rochester NY

Investigators

Abstract

The proposed research program consists of experimental and computational studies of human visual learning. The project focuses on the information processing mechanisms mediating the perceptual learning that underlies expertise in a variety of STEM fields, such as biology, astronomy, and geoscience. In particular, the investigators attempt to take advantage of insights from the field of Machine Learning (e.g., its formalisms for conceptualizing the properties of different learning environments, its powerful sets of statistical learning algorithms for each environment, and its numerous mathematical and empirical findings about the advantages and disadvantages of these algorithms). The studies look at learning performance on lower-level and higher-level discrimination tasks in four types of learning environments: supervised, unsupervised, semi-supervised, and reinforcement learning environments. The project also explores visual learning based on correlated perceptual signals in multisensory or multi-cue environments, such as when a person both sees and touches surfaces. The computational studies compare people's learning performances with the statistically optimal performances of "ideal learners", and also with the performances of on-line learning algorithms from the Machine Learning literature. A key hypothesis is that people can visually learn with "unlabeled" data items (i.e., items that are not labeled by an instructor as examples of a particular category of interest) by transferring knowledge gained with "labeled" data items or by transferring knowledge gained from other sensory modalities. The work has important implications for the design of STEM training environments.