Efficient source detection using integrate-and-fire neurons

Sensory data extracted by neurons is often noisy or ambiguous and a goal of low-level cortical areas is to build an efficient strategy extracting the relevant information. It is believed that this is implemented in cortical areas by elementary inferential computations dynamically extracting the most likely parameters corresponding to the sensory signal. We explore here a neuro-mimetic model of the feed-forward connections in the primary visual area (VI) solving this problem in the case where the signal may be idealized by a linear generative model using an over-complete dictionary of primitives. Relying on an efficiency criterion, we derive an algorithm as an approximate solution which provides a distributed probabilistic representation of input features and uses incremental greedy inference processes. This algorithm is similar, to Matching Pursuit and mimics the parallel and event-based nature of neural computations. We show a simple implementation using a network of integrate-and-fire neurons using fast lateral interactions which transforms an analog signal into a list of spikes. Though simplistic, numerical simulations show that this Sparse Spike Coding strategy provides an efficient representation of natural images compared to classical computational methods.