What do cognitive models and human judgments suggest about the desired structure of automatic classifiers?

Abstract

Past studies probed human listeners’ efficacy at classification of impulsive sonar echoes by using paired‐comparison ratings to measure perceptual dissimilarity. Interpreting these ratings requires a cognitive model comprising both a representation of the stimuli and a process operating on that representation. Initially, perceived dissimilarities were represented as distances in Euclidean space via multidimensional scaling. This assumed a continuous and spatial cognitive representation and proved difficult to relate to a linear vector space of features. Later work by the authors suggested a discrete (categorical) cognitive representation may better reflect perception of these stimuli. While similarity‐based classifiers can bypasses feature extraction [S. Philips and J. Pitton, J. Acoust. Soc. Am. 123, 3344 (2008) (A)], the form of the similarity measure reflects the assumed cognitive representation [L. Cazzanti and M. R. Gupta, Proc. IEEE Intl. Symposium Info. Theory, pp. 1836–1849 (2006)]. Here, findings based on paired‐comparison ratings and verbal descriptions of stimuli are discussed in terms of the cognitive models they reflect and implications for classifier architecture: the form of similarity measures, whether features are continuous or discrete, and whether features are common to all classes. [Work supported by the Office of Naval Research.]