A simulation framework for arti cial neural systems

A simulation framework for articial neural network models and electronic implementations (CMOS) is presented. Neural network models spanning from arti-cial biological models through mathematical or computational models such as back-propagation type networks to digital, analog or subthreshold analog implementations can be simulated concurrently. Neural and electronic variables (nodes) interact through simple mapping functions, thus allowing a stepwise electronic implementation of an articial neural network system. Nonlinear devices (e.g. MOS FET transistors and neurons) are modelled by table look-up methods using cubic Bspline interpolation to ensure continuous rst derivatives. This table look-up method provides fast and accurate models with small tables. Key features of the simulation framework are mixed mode simulation for variable accuracy/speed trade o, event-driven approach exploiting latency and multirate behavior for simulation speed-up. Multilevel abstraction and macro modelling of current mode analog subthreshold circuits are provided. Automatic switching from macro level to transistor level guarantees reliable simulation.