Neurobiological Inspiration for the Architecture and Functioning of Cooperating Neural Networks

In order to emulate more complex and more realistic human-like functions, it is now well admitted that a single monolithic neural network is not sufficient. Biological data show that the cortex is a set of interconnected neural networks. Beyond the classical view of one way feedforward neural network guiding an information flow from an input to an output layer, we now have to imagine architectures and functioning rules that permit cooperation and information exchange between such neural networks. Inspired with biological data, we propose here such a scheme bringing into play of complex units like the cortical column, a functional micro-circuit repeated throughout the cortex. This basic unit of treatment gathers the classical weighted sum for feedforward information flow and sigma-pi operations for cooperation between different axis of treatment. We illustrate this model with an application of cooperation between character recognition and localization.