Simulation of artificial neural networks with biological features based on building blocks

Exact modeling and fast simulation of neural networks using conventional von-Neumann computers present a great challenge. Due to inherent parallelism of such networks extreme computational efficiency is required in order to obtain reasonable CPU times. In this contribution we use neurons transmitting neural information by firing pulse trains, that is, in the same biological forms as in the nature. The resulting mixed digital/analog design of such artificial neural networks is a great complexity. Therefore, the simulation which uses conventional circuit simulation for the analog cirucit parts is not practicable. We present a library of parameterized building-blocks for artificial neurons with biological features. This library allows efficient simulation and analysis of neural networks using the hierarchical mixed-mode simulator DOSCA. Some typical examples of simulated-training-phases and neural computing are presented.