Addressing scaling and plasticity problems with a biologically motivated self-organizing network

The scaling and plasticity problems present two major obstacles to the development of large-scale neural network (NN) systems. Through evolution, biological networks have solved these problems to create systems which dwarf even the largest artificial attempts. Biological evidence supports an architecture built upon clusters of nodes (neurons) which are combined to form insulated modules called neural processing units (NPUs). Two types of connectivity distinguish this architecture. Connection between NPU modules is facilitated by a hardwired technique called gross connectivity, whereas connections between nodes within a NPU are defined by a plastic, self-organizing technique called fine connectivity. Fine connectivity is dependent on three main elements: the inherent characteristics of the node, the local environment of the node, and the specific input modality which innervates an NPU. It is demonstrated that this architecture significantly resolves the scaling and plasticity problems for large networks