Domain knowledge and feature representation

Abstract

Identifying covert internal brain by their expression in neural images, particularly from magnetic resonance imaging, is a popular, powerful, and important area of research whose ultimate expression is known now as “brain reading.” The nature of the imaging data is challenging however, in that they typically have two orders of magnitude more features than observations. We propose that it is many ways useful to apply prior knowledge of brain organization - both physical and mental. We note that sparsity offers quantitative leverage, and that this, itself, may provide insight into the nature of human cognition.