Neural coding in graphs of bidirectional associative memories

Abstract In the last years we have developed large neural network models for the realization of complex cognitive tasks in a neural network architecture that resembles the network of the cerebral cortex. We have used networks of several cortical modules that contain two populations of neurons (one excitatory, one inhibitory). The excitatory populations in these so-called “cortical networks” are organized as a graph of Bidirectional Associative Memories (BAMs), where edges of the graph correspond to BAMs connecting two neural modules and nodes of the graph correspond to excitatory populations with associative feedback connections (and inhibitory interneurons). The neural code in each of these modules consists essentially of the firing pattern of the excitatory population, where mainly it is the subset of active neurons that codes the contents to be represented. The overall activity can be used to distinguish different properties of the patterns that are represented which we need to distinguish and control when performing complex tasks like language understanding with these cortical networks. The most important pattern properties or situations are: exactly fitting or matching input, incomplete information or partially matching pattern, superposition of several patterns, conflicting information, and new information that is to be learned. We show simple simulations of these situations in one area or module and discuss how to distinguish these situations based on the overall internal activation of the module. This article is part of a Special Issue entitled “Neural Coding”.