The Development of a Research Environment for Neural Networks: Instantiating Neocognitions
GEORGIA TECH RESEARCH INST ATLANTA ARTIFICIAL INTELLIGENCE BRANCH
Pagination or Media Count:
Neural networks can be thought of as combinations of generic pieces linked together in varying architectures. Many different models and architectures have been presented in the published literature. Networks may differ both in the characterization of their pieces and in the connection patterns of those pieces. In order to exploit the similarities between models, incorporate the differences between models, and automate the process of linking the pieces together, a prototype of a generalized research environment for neural networks is being developed. The main virtue of this generalized environment is the flexibility it provides for testing neural network architectural and processing decisions without having to write programs. The environment encompasses the ability to specify desired characteristics e.g., activation functions, connection masks, sub-net sizes as parameters to network creation functions it does not force a programmer to combine such characteristics by altering the program code itself. This paper initially introduces the generalized research environment, subsequently discusses the architecture of a test case network the neocognitron, and finally presents the initial results in testing a neocognitron instantiated by the environment.