Perceptual Sensitivity is Explained by Optimization for Ecological Hearing Tasks 

Background: When asked to discriminate between two stimuli that differ along some dimension, such as intensity, frequency, or orientation, humans exhibit a smallest stimulus difference, or threshold, below which discrimination becomes impossible. These limits on perceptual sensitivity have traditionally been explained by internal noise. Contemporary theories instead link perceptual sensitivity to efficient coding of the natural environment but have not been comprehensively tested. Here, we measured the psychoacoustic sensitivity of a computational model optimized for everyday hearing tasks.

Methods: We first optimized a deep artificial neural network to jointly localize and recognize words, voices, and environmental sounds using simulated auditory nerve representations as input. We then estimated the model’s psychoacoustic thresholds by training linear classifiers to make simple discrimination judgments using the network’s learned representations.

Results: Thresholds derived from ecologically task-optimized representations exhibited qualitative and quantitative matches to human sensitivity across a large set of perceptual dimensions. By contrast, thresholds derived from alternate model representations did not match those of humans. Untrained and peripheral representations tended to underestimate human sensitivity, whereas representations specifically optimized for psychoacoustic tasks tended to produce superhuman sensitivity.

Conclusions: The results show that many classical perceptual phenomena (such as spectral/temporal modulation sensitivity and binaural unmasking) are side effects of representations optimized for natural listening behavior, and raise the possibility that psychophysical thresholds are determined by linear separability in these optimized representations, rather than intrinsic neural noise.