Spiking network model of A1 learns temporal filters with frequency preferences

Abstract

AbstractThe sparse coding hypothesis has successfully predicted neural response properties of several sensory brain areas. For example, sparse basis representations of natural images match edge-detecting receptive fields observed in simple cells of primary visual cortex (V1), and sparse representations of natural sounds mimic auditory nerve waveforms. SAILnet, a leaky integrate-and-fire network model (“Sparse and Independently Local network”) has previously been shown to learn simple V1 receptive fields when trained on natural images. Experimental work rewiring visual input to auditory cortex found that auditory neurons developed visual response properties, suggesting that developmental rules may be shared across sensory cortices.In this work we adapt SAILnet to train it on waveforms of auditory sounds and learn temporal receptive fields (filters), in contrast with previous work that trained SAILnet or other network models on spectrograms. In our model network of primary auditory cortex (A1) neurons receive synaptic current from input neurons who temporally filter the direct sound waveforms. To show the network learns frequency-dependent filters naturally, we do not parametrize the temporal filters, and only restrict the total number of time points in the filters. To make training feasible, we simplify the model to a single input neuron and 768 A1 neurons, and we train the network on “lo-fi” music, whose spectral power is limited to frequencies of∼10, 000 Hz or less, giving a manageable temporal resolution of the stimulus and filters. The learned filters develop distinct frequency preferences, and reconstruction of novel stimuli captures the low-frequency content of signals in reasonable detail, with audio playback capturing clear aspects of the original stimulus. Lastly, our work also has a pedagogical benefit: the learned stimulus features can be played as sounds, which aids in teaching sensory coding to learners with visual impairments who cannot perceive stimulus features learned by V1 models.