Inferring Image Templates from Classification Decisions

Arnab Dhua, Florin Cutzu

Abstract: Assuming human image classification decisions are based on estimating the degree of match between a small number of stored internal templates and certain regions of the input images, we present an algorithm which infers observers classification templates from their classification decisions on a set of test images. The problem is formulated as learning prototypes from labeled data under an adjustable, prototype-specific elliptical metric. The matrix of the elliptical metric indicates the pixels that the template responds to. The model was applied to human psychophysical data collected in a simple image classification experiment.