Efficient coding of correlated complex acoustic dimensions through active listening.

Abstract

Efficient coding extracts and exploits redundancy to optimize information processing in sensorineural systems. Stilp et al. [J. Acoust. Soc. Am. 124, 2496 (2008)] reported evidence for rapid and efficient adaptation to correlation among complex acoustic attributes. In the study, following passive exposure to highly correlated stimulus features, discriminability of sound pairs violating the correlation is temporarily lost before subsequent recovery via active testing with stimuli whose features were poorly correlated. The present study examines listeners’ ability to extract and exploit correlation between stimulus attributes exclusively through active testing. Listeners discriminated stimuli (AXB) for which two complex, independent dimensions, attack/decay (AD) and spectral shape (SS), were highly correlated (r2=0.96). In the first testing block of 128 trials, discrimination of sound pairs respecting the correlation is superior to sound pairs violating the correlation. Only through successive testing blocks, listeners discover variance orthogonal to the otherwise perfect correlation between AD and SS, and discrimination recovers to baseline levels. Listener performance will be discussed within the context of maximum likelihood and connectionist models. [Work supported by NIDCD.]