Adaptive efficient coding of correlated acoustic properties

Abstract

word)Natural sounds such as vocalizations often have co-varying acoustic attributes where one acoustic feature can be predicted from another, resulting in redundancy in neural coding. It has been proposed that sensory systems are able to detect such covariation and adapt to reduce redundancy, leading to more efficient neural coding. Results of recent psychoacoustic studies suggest that, following passive exposure to sounds in which temporal and spectral attributes covaried in a correlated fashion, the auditory system adapts to efficiently encode the two co-varying dimensions as a single dimension, at the cost of lost sensitivity to the orthogonal dimension. Here we explore the neural basis of this psychophysical phenomenon by recording single-unit responses from primary auditory cortex (A1) in awake ferrets exposed passively to stimuli with two correlated attributes in the temporal and spectral domain similar to that utilized in the psychoacoustic experiments. We found that: (1) the signal-to-noise (SNR) ratio of spike rate coding of cortical responses driven by sounds with correlated attributes was reduced along the orthogonal dimension; while the SNR ratio remained intact along the exposure dimension; (2) Mutual information of spike temporal coding increased only along the exposure dimension; (3) correlation between neurons tuned to the two covarying attributes decreased after exposure; (4) these exposure effects still occurred if sounds were correlated along two acoustic dimensions, but varied randomly along a third dimension. These neurophysiological results are consistent with the Efficient Learning Hypothesis and may deepen our understanding of how the auditory system represents acoustic regularities and covariance. Significance (119 words)In the Efficient Coding (EC) hypothesis, proposed by Barlow in 1961, the neural code in sensory systems efficiently encodes natural stimuli by minimizing the number of spikes to transmit a sensory signal. Results of recent psychoacoustic studies are consistent with the EC hypothesis, showing that following passive exposure to stimuli with correlated attributes, the auditory system adapts so as to more efficiently encode the two co-varying dimensions as a single dimension. In the current neurophysiological experiments, using a similar stimulus design and experimental paradigm to the psychoacoustic studies of Stilp and colleagues (2010, 2011, 2012, 2016), we recorded responses from single neurons in the auditory cortex of the awake ferret, showing adaptive efficient neural coding of correlated acoustic properties.

Long-standing problems in speech perception dissolve within an information-theoretic perspective

“…While the neural locus or loci responsible for these human perceptual findings remains an open question, multiple neurophysiological studies with animal models are instructive. Most recently, findings by Stilp and colleagues have been replicated in neural response in primary auditory cortex of ferrets upon presentation of sounds that covaried in amplitude modulation and peak frequency of the spectral envelope (Lu, Liu, Dutta, Fritz, & Shamma, 2019). Following exposure to stimuli capturing these correlated attributes, signal-to-noise ratio (SNR) of spike-rate coding decreased orthogonal to the correlation, while remaining intact along the correlation; and, mutual information of spike coding increased only along the correlated dimension.…”

Long-term priors constrain category learning in the context of short-term statistical regularities

“…If underlying representations change as a consequence of short-term passive exposure to regularities as suggested in prior research (Lu et al, 2019;Stilp & Kluender, 2012;Stilp et al, 2010) one might predict that this exposure also would impact how participants use acoustic information learn novel categories. The basis of this expectation stems from a substantial literature demonstrating that the input dimensions across which category exemplars are defined and the sampling of exemplars from an input space are crucial in predicting how category learning proceeds (Ashby & Maddox, 2011;Folstein, Gauthier, & Palmeri, 2010;Gureckis & Goldstone, 2008;Maye et al, 2002;Smith, Jayaraman, Clerkin, & Yu, 2018;Werker, Yeung, & Yoshida, 2012).…”

“…These prior investigations (Stilp & Kluender, 2012;Stilp et al, 2010) have focused on one kind of relationship (positive or negative correlation) in one acoustic environment (attack-decay and spectral shape) and examined behavior using discrimination measures. A recent study found that neurons in ferret auditory cortex reflect rapid efficient coding of correlations of similar acoustic attributes (peak frequency of the spectral envelope and amplitude modulation; Lu et al, 2019). These prior studies highlight that sensory systems are able to rapidly adapt to novel regularities reflecting correlations between acoustic attributes changing both discrimination behavior and neural tuning in auditory cortex.…”

“…Some studies have made a case that statistical learning across short-term regularities may also be 'efficient.' Examples of 'rapid efficient coding' demonstrate the rapid learning of novel input regularities, even in a passive manner (Barlow & Földiák, 1989;Coen-Cagli, Kohn, & Schwartz, 2015;Lu, et al, 2019;Stilp, Rogers, & Kluender, 2010).…”

“…However, prior research on rapid efficient coding in sensory systems suggests that the mechanisms underlying this rapid adaptation can be accomplished through passive exposure (Lu et al, 2019;Stilp, Kiefte, & Kluender, 2018;Stilp et al, 2010;Stilp & Kluender, 2012. It could be the case that to see broader impacts or changes to representations, more active engagement or feedback may need to be present.…”