Speaker
Description
Complex stimuli are represented by the activations of populations of neurons in the visual cortex. While the response pattern of a single neuron provides limited stimulus information, the collection of the responses from a population can be used to reliably identify the image eliciting the responses. A key question regarding the neural code is whether the information carried by a population is simply the sum of the individual neural contributions or joint activation patterns provide further information. Recent studies demonstrated that a linear decoder, which does not take into account joint activations, is sufficient to decode information about the orientation of grating images. Representation of more complex images, however, is expected to exploit knowledge about statistical regularities in the presence of constituent features and therefore such statistical regularities can shape the activation patterns of neurons representing those features. We constructed decoders that capture different statistical structure in multiunit responses recorded from the primary visual cortex of macaques. We used natural and synthetic images to investigate how the presence of statistical structure in responses affects decoding performance. We learned that adding quadratic features to logistic regression, which capture joint activations of pairs of neurons, did not enhance decoding performance of properly fitted decoders. We constructed mixture decoders, which could characterize the contribution of stimulus-dependent correlations or stimulus-dependent variances to nonlinear effects. Using mixture decoders we demonstrated that stimulus-dependent spike count correlation structure contributes to nonlinear decoding capabilities. Finally, comparing the coding of complex natural image patches and that of limited-complexity synthetic images we showed that a nonlinear decoding strategy is more advantageous for complex images than for images with simpler structure. Taken together, our results highlight that structure in stimuli introduces intricate joint statistics in V1 responses, which has the consequence that stimulus identity can be most efficiently established with nonlinear decoders.
