ABSTRACT
It is argued that the state-of-the-art for design and simulation tools for NN architectures is still not satisfactory. The approach presented in this paper is a very general block diagram approach formalized as data flow graphs. A prototype called SESAME 1 [10] has been implemented in C++. We show that SESAME improves the state-of-the-art in neurosimulation considerably. It allows elegant and extensible construction of multiple neural net paradigms and combinations thereof. The basic framework includes run-time parameterized multiple inheritance, nested parametrization and automated checks of wellformedness, which in practice turn out to be remarkably useful. For illustration it is shown how to specify arbitrary back propagation architectures. The presented framework conforms to the latest findings of software engineering with respect of reuse and extensibility [5].
