Inferring the neural basis of binaural phenomena with a modified autoencoder
May, 2021
About
Broad organizational features of the auditory pathway have been replicated by artificial neural networks trained on diverse and naturalistic stimuli. It is credible that neural network analogues could also reveal specific neural computations underlying targeted auditory phenomena. This remains to be established. Here, we examined a modified autoencoder (a deep neural network akin to symbolic regression) trained to imitate an auditory behavior rooted in the binaural system. The behavior promoted the use of split-second interaural timing cues to improve the detection of a tone presented in noise. In the optimal network, we observed the emergence of specialized computations with prominent similarities to animal models. Artificial neurons developed a sensitivity to temporal delays that increased deeper into the network and were widely distributed in preference (extending to delays beyond the range permitted by head width), and the ensuing dynamics were consistent with a binaural cross-correlation mechanism. Our results attest to the generality of these solutions for performing signal detection at low frequencies. Moreover, this is a primary demonstration that deep learning methods can be used to infer tangible mechanisms underlying auditory perception.