Recent progress on reading the retinal code
Decoding of complex stimuli from the retinal activity remains an open challenge. To date, experiments have focused on decoding either a small number of discrete stimuli or very low dimensional dynamical traces. We stimulated a rat retina with a rich new class of synthetic stimuli, in which small circular spots execute random dynamical motion. We constructed linear and nonlinear decoders that reconstruct the full stimulus movie from the simultaneously recorded activity of ~100 retinal ganglion cells. Linear decoders reveal a local and sparse structure of the code with a low amount of redundancy, in apparent contrast to previous work that used simpler stimuli. Nonlinear decoding substantially improves linear decoding performance. We show that nonlinear decoding makes use of higher order statistics of the spike trains to suppress decoder output when the cells spike spontaneously or when there are no stimulus changes in their receptive field centers. These statistics are strongly affected by temporal and across-cell noise correlations. We hypothesize that our results are applicable more generally to the problem faced by the brain, where downstream areas need to discriminate whether the incoming signals are due to stimuli or spontaneous activity, and suggest that noise correlations could play a central role in decoding.