Modeling processing of complex sounds by neurons at the cortical level

Cortical auditory processing is critical for speech perception, yet we understand little about how the neural representations available at the cortical level allow for sound recognition. In particular, the neural mechanisms for recognition of sounds via spiking patterns are currently unknown, and the characterization of auditory cortical neuron response properties using stimulus-response models remains incomplete.;Previous work has shown that spike trains can be used to discriminate between sounds using analytical methods, but how biological circuits might perform such computations remains unclear. Therefore we devised a model for discrimination (inspired by a spike distance metric) that combines a network of integrate-and-fire model neurons with a decision network. We applied this model to the birdsong system, a powerful animal model for the study of audition.;We then examined how stimuli could be recognized by sensory systems from spiking patterns by transforming spiking patterns into spatial representations. Previous "spike pattern recognition" models used artificially pre-processed and noiseless signals as input, despite the fact that neural spike trains show variability. Therefore, we used cortical-level neural recordings as input to a novel spike pattern recognition system designed to deal with the intrinsic variability and diverse response properties of cortical spike trains. We show that the model can learn to recognize neural responses to auditory stimuli using spike-timing dependent plasticity, and can play back learned spike patterns in reverse, not unlike the reverse spike train playback observed in hippocampus.;A standard model for central auditory neurons is the spectro-temporal receptive field (STRF), which quantifies the spectral and temporal stimulus parameters that modulate neural activity. While the STRF has been used to predict neural responses, a causal connection between the STRF structure and neural responses is lacking. We used an adaptive stimulation paradigm to address this problem. Specifically, for each neuron, we filtered out stimulus frequencies predicted to be unimportant by the STRF (estimated using normalized reverse correlation), and measured responses to these stimuli. For some neurons, this filtering had little effect on neural firing, but adding a noise masker in the unimportant frequency regions adversely affected their responses, suggesting that frequency regions outside the STRF can play a role in determining neural responses.