If modern computers are sometimes superior to humans in some specializedtasks such as playing chess or browsing a large database, they can't beat theefficiency of biological vision for such simple tasks as recognizing andfollowing an object in a complex cluttered background. We present in this paperour attempt at outlining the dynamical, parallel and event-based representationfor vision in the architecture of the central nervous system. We willillustrate this on static natural images by showing that in a signal matchingframework, a L/LN (linear/non-linear) cascade may efficiently transform asensory signal into a neural spiking signal and we will apply this framework toa model retina. However, this code gets redundant when using an over-completebasis as is necessary for modeling the primary visual cortex: we thereforeoptimize the efficiency cost by increasing the sparseness of the code. This isimplemented by propagating and canceling redundant information using lateralinteractions. We compare the efficiency of this representation in terms ofcompression as the reconstruction quality as a function of the coding length.This will correspond to a modification of the Matching Pursuit algorithm wherethe ArgMax function is optimized for competition, or Competition OptimizedMatching Pursuit (COMP). We will in particular focus on bridging neuroscienceand image processing and on the advantages of such an interdisciplinaryapproach.
展开▼
