How should we characterize neural responses to natural stimuli? To address this question, neuroscientists typically look at either of two complementary aspects of neural representations. The first, and best studied, is the encoding process by which a stimulus is converted by the nervous system into neural activity. Less commonly studied is the decoding process, by which experimenters attempt to use neural activity to reconstruct the stimulus that evoked it. Several studies have suggested an asymmetry in these processes: encoding tends to be nonlinear, whereas decoding tends to be linear. However, such an asymmetry has never been shown directly in the same neurons.
Here we use information theoretic techniques to directly compare encoding and decoding processes in auditory cortical cells responding to natural stimuli. As we expected, information captured by a linear encoding (STRF) model is far less than the total information. Contrary to our expectation, however, a linear decoding model captures as little information as a linear encoding model, suggesting that the performance of a linear model is equally bad for both encoding and decoding processes.
Poster (167KB, PDF):
Computational and Systems Neuroscience (COSYNE) 2005, Salt Lake City, Utah
Methods: Technical Notes on Linear Regression and Information Theory
See also my dissertation (Appendix A).