Modeling and characterization of neural gain control
Sensory neurons often exhibit striking nonlinear behaviors that are not adequately described by a linear receptive field representation.
In the first part of the work, we suggest that these nonlinearities arise because sensory systems are designed to efficiently represent environmental information. We describe a form of nonlinear decomposition (specifically, divisive gain control) that is well-suited for efficient encoding of natural signals. We show that this decomposition, with parameters optimized for the statistics of a generic ensemble of natural images or sounds, can account for some nonlinear response properties of “typical” neurons in both vision (area V1) and audition (auditory nerve). This work provides theoretical justification to neural models of gain control, and explains how one might choose the parameters of the model based on efficient coding considerations.
In the second part of the work, we describe a methodology for characterizing this class of nonlinear sensory models. The characterization is based on a white noise analysis, in which a set of random stimuli are presented to a neuron and the spike-triggered ensemble (specifically, the spike-triggered covariance) is analyzed. We demonstrate the applicability of the technique to retinal ganglion cell data in monkey and salamander.