How Can a Neuron Interpret a Spike Train?

Michael Wu, University of California, Berkeley

Spike trains are commonly thought to be stochastic and noisy. For this reason neurophysiologists often compute the peristimulus time histogram (PSTH) by averaging responses over repeated stimulus presentations. During natural vision, however, a neuron produces only a single spike train in response to each stimulus. Downstream neurons do not have access to the time-averaged responses of upstream cells. Given this limitation, how well can a neuron interpret a single spike train?

To address this issue we recorded spike trains from single V1 neurons in awake macaques during repeated presentations of movies simulating natural vision. Individual spike trains are poor predictors of the observed PSTH (r=0.38, n=60, 14 ms bins). However, different decoding strategies can significantly improve PSTH prediction. When the optimal linear filter is used to map spikes into the PSTH, predictions improve substantially (r=0.51). A more sophisticated method is to project each spike train into the principal component (PC) subspace that maximizes signal variance after removing noise PCs. This method further increases the correlation between the predicted and observed PSTH (r=0.58). An alternative decoder is based on independent component (IC) analysis. Because ICs are temporally localized and statistically independent, they may correspond to specific biophysical mechanisms. When individual spike trains are projected into the IC subspace, the prediction correlation is similar to that obtained with PCA (r=0.59). These methods for decoding spike trains may be particularly valuable for interpreting results of single-trial experiments (such as free-viewing visual search), where average response data are not available.

Abstract Author(s): Michael CK Wu; Stephen V David; Jack L Gallant