Human and animal brains employ many variations on common underlying neural structures to achieve a range of abilities. These neural structures are used by all five senses and give us the ability to control our bodies, to communicate, to think, and to feel emotions. At the systems level, these structures have been characterized as recurrent neural networks. Recurrent neural networks share a general mathematical formalism using non-linear differential equations. This formalism is quite specific to biologically plausible recurrent neural networks. It encompasses, for example, neural network models of vision, memory, motor control, behavioral conditioning, and many others. Though this common mathematical formalism exists, there exists no general description language to rapidly specify and implement these models on the computer. Development of sophisticated architectures requires highly trained personnel and is very time consuming. Often, it is not economical to develop a custom architecture specific to a given problem area. We propose to address this problem with a two-component solution. The enclosed proposal details the first component of that solution: It proposes a computer interpretable language that can represent a host of neurobiologically plausible neural network models. Note that this is not 'just another neural network simulator': Unlike most artificial neural networks, recurrent neural networks actually model the brain.
Keywords: RECURRENT NEURAL NETWORKS COGNITIVE & NEURAL SCIENCES MASSIVELY PARALLEL COMPUTING GENETIC ALGORI
