A simple reverse correlation scheme for the identification of visual neurons

Purpose. The standard approach to generalize the white-noise technique to neural systems with multiple inputs consists of using a spatio-temporal white noise stimulus. A drawback of this methodology is that the input space to be explored is huge, and only a sparse coverage can be achieved in limited time. We propose a new discrete-time reverse correlation technique that effectively reduces the dimension of the input space, yielding higher signal to noise ratios. This is achieved by exploiting a priori knowledge about the spatial tuning properties of the neuron. Results. We first select a set S of M orthonormal images of size N2 pixels. The idea is to have M ≪ N2 and use previous knowledge about the neuron's spatial tuning to select an appropriate input space. An input image sequence is generated by selecting, at each time, a random element from S. We prove that the projection of the receptive field onto the subspace spanned by the set S can be estimated based on measurements of the crosscorrelation between the input image sequence and the cell's output. The technique can also be applied to systems that can be modeled as a linear receptive field followed by a static nonlinearity. Examples are shown where S is a subset of the complete two-dimensional discrete Hartley basis functions. Conclusions. A simple reverse correlation scheme that only requires the generation of a fixed number of images can be used to identify quasi-linear visual neurons. Prior knowledge of the spatial tuning of the cell can be incorporated in the selection of an effective set of stimulus images. We are currently applying this technique to the analysis of V1 simple cells.

Duke Scholars

Author Guillermo Sapiro Electrical and Computer Engineering