Classification of impulsive-source active-sonar echoes based on a model of auditory perception

Abstract

Models of auditory perception are investigated as the basis for a biomimetic classifier of impulsive-source active-sonar echoes. Multidimensional scaling estimates the perceptual space in which listeners perform classification [J. E. Summers et al., J. Acoust. Soc. Am. 120, 3125 (A) (2006)]. In the resulting space of perceptual dimensions, stimuli form distinct clusters and target is discriminated from clutter along a single perceptual dimension. Dimensions in this space do not correspond to features having simple algorithmic representations. Consequently, conventional methods to develop a mapping from signal space to feature space fail. Instead, dimensions reflect untrained categorical perception manifested through the mixtures of top-down and bottom-up processes used by listeners: A high-level cognitive process for interclass dissimilarity ratings and a low-level signal-based process for intraclass comparisons. Behaving as expert systems, listeners rapidly assign stimuli to categories based on prior experiences, a process analogous to the statistical description of classes in the class-specific method. In contrast, within-class comparisons reflect signal-derived features found to be most efficacious for differentiating between the signals, a process similar to generation of features by singular-value decomposition. Implications of these findings for design of hybrid generative/discriminative human-mimetic classifier architectures are discussed. [Work supported by ONR.]