Feedforward inhibition and inhibitory synaptic plasticity generate sparse, selective, and background-invariant representations of auditory stimuli in a spiking model of zebra finch caudomedial nidopallium

Pattern decorrelation and the separation of signal from noise are defining aspects of several sensory circuits. In a recent study of the zebra finch auditory stream, Schneider and Woolley (2013) reported a population of higher-order auditory cortex cells—broad spiking units of the caudomedial nidopallium (NCM)—that responded to specific sets of zebra finch songs, exclusively when the songs were played at intensities that permitted behavioural recognition against background noise. Here we describe a spiking model of the broad-spiking NCM cells (and accompanying circuit) that accurately replicates the sparse firing rates, selectivity patterns, and signal-to-noise dependences observed in vivo. Using leaky integrate-and-fire neurons as outputs and the physiologically recorded primary auditory cortex trains from Schneider and Woolley (2013) as templates for input ensembles, we demonstrate that the implementation of inhibitory synaptic plasticity in a feedforward inhibitory module with a dynamic spiking threshold is sufficient to produce the selective, non-linear receptive fields necessary for cell-song specificity and behaviourally relevant firing patterns. Notably, the model recreates a linear increase in output firing rate as a function of the signal-to-noise ratio—but not of the input firing rate—in a manner that is consistent with both the biologically recorded NCM firing patterns and the behavioural variables relevant to signal extraction. We also describe preliminary evidence regarding multiple impinging signals (songs) and the embedding of these feedforward modules in larger networks with lateral connectivity.