A scalable Khoros neural network toolbox

Abstract

Three problems limit neural network simulation environments. First, the computational complexity of neural network simulation imposes a practical constraint on the size of simulated neural networks. Second, neural network simulation environments are usually designed to accommodate a confined set of neural network models. Third, most neural network simulation environments do not facilitate interoperability with external systems. This project addresses these issues by implementing a simulation environment that scales to support larger neural systems, facilitates the implementation of new network architectures, and interoperates with other software. This problem was approached by examining a representative set of neural network architectures that feature a diversity of characteristics such as topology and learning rules. The result of this research was applied to the design and implementation of an environment consisting of tools and applications to support neural network simulation. This environment, developed under the Khoros system, provides a visual neural network construction tool, an extensible C++ class library that encapsulates network management and object interaction, a file format for storage and retrieval of neural networks, and finally, neural network visualization tools.