Automated model-based spike-sorting: a method for classification with confidence intervals
New York University
Spike-triggered average (reverse correlation) techniques are effective for linear characterization of neural responses. But sensory neurons exhibit important nonlinear behaviors that are not captured by such analyses. Many of these nonlinear behaviors are consistent with gain control. We assume a model in which the spike rate of a neuron is proportional to the halfwave rectified response of a linear kernel suppressively modulated by a weighted sum of rectified responses of other linear kernels. We develop a spike-triggered covariance method for recovering the suppressive axes of such a model. First, we recover a linear kernel through spike triggered averaging. Next, we compute the covariance matrix of the stimuli eliciting spikes, and perform a principal components decomposition of this matrix. The principal axes (eigenvectors) associated with small variance (eigenvalue) correspond to directions in which the response of the neuron is modulated suppressively. Finally, we recover the suppressive weights associated with these axes by maximizing the likelihood of the spike data. We demonstrate this method on simulated examples and on salamander retinal ganglion cell data (from the Chichilnisky lab). Preliminary analysis of the physiological data reveals meaningful suppressive axes and explains interesting nonlinearities. We believe this method will be applicable to other sensory areas and modalities.