Attribute grammar encoding based upon a generic neural markup language: facilitating the design of theoretical neural network models

Abstract

There is a need for tools that facilitate the systematic exploration of novel theoretical neural network models. Existing neural network simulation environments, neural network specification languages, and genetic encoding of neural networks fall short of providing the tools needed for this task. We suggest that a useful approach to the design of such tools may be the use of a grammar to capture neural design principles as well as structural and behavioral elements, and mechanisms to automatically translate the parse trees of the grammar into complete neural network specifications in a generic format. We present the attribute grammar encoding (AGE) as a specific example of our approach that uses attribute grammars to create descriptions of neural network solutions in an XML-based format termed the generic neural markup language (GNML). Lessons learned from the development of this system are presented to identify and address the issues of a broader application of this approach to other specification formats and other grammar encoding approaches.