Optimal signal representation in neural spiking population codes as a model for the formation of simple cell receptive fields.
Résumé
The primary visual cortex is the central hub for the transmission of visual information to the rest of the central nervous system. We study here models explaining how this neural system becomes efficient through natural selection and neural development. By defining and then optimizing an efficiency cost, we may derive an adaptive model of the input to the primary visual cortex as an unsupervised learning algorithm. As an alternative to classical models, we focus here on the fact that visual information is carried from the sensory organs to the primary visual cortex by neuronal events, or spikes, in bundles of parallel fibers. In fact, taking advantage of the constraint that spikes may be considered as all-or-none binary events, we may build a generic cost for the efficiency of the visual representation as a measure of the L$_0$ norm sparseness of the code. However, this is a ``hard'' NP-complete problem and we propose a solution in a population of generic Integrate-and-Fire neurons. It relies both on a correlation-based inhibition using lateral interactions and on an homeostatic constraint by a spiking gain control mechanism, two key features of cortical processing. For comparison purposes, we applied this scheme to the learning of small patches taken from natural images and compared the results and efficiency with state-of-the-art algorithms. Results show that while the different coding algorithms gave similar efficiencies, the homeostasis provided an optimal balance which was crucial during the learning. This study provides a simpler yet more efficient algorithm for learning that is particularly well adapted to the architecture of neural computations. By providing the optimally independent components in a set of inputs, it suggests that this Sparse Spike Coding strategy may provide a generic computational module.
Origine | Fichiers produits par l'(les) auteur(s) |
---|