Rapid efficient encoding of correlated complex acoustic properties.

Abstract

Early sensory processing is long thought to extract and exploit redundancy and regularity in the input, but perceptual evidence supporting this hypothesis has been limited. Experiments are reported demonstrating efficient encoding of multiple acoustic attributes resulting in the reduction of perceptual dimensionality. Stimuli varied across two complex independent dimensions: attack/decay (AD) and spectral shape (SS). A subset of 24 stimuli captured a near-perfect correlation between these dimensions (r2= 0.95). Participants passively listened to 25 presentations of these correlated stimuli over 7.5 min before discriminating sound pairs in AXB trials without feedback. For trials where both SS and AD varied, listeners successfully discriminated sounds consistent with the correlation of SS and AD during exposure, but they were at chance discriminating pairs varying orthogonal to the correlation presented in exposure. After exposure, participants were also significantly poorer at detecting differences among stimuli varying in only SS or AD. Thus, passive listening to stimuli with correlated attributes resulted in the collapse of two physically and perceptually distinct stimulus dimensions onto a single dimension that efficiently encoded exposure covariance. Data are consistent with Hebbian models of cortical adaptation but contradict models derived from anti-Hebbian and competitive-learning principles. [Work supported by NIDCD.]