Realization of a neural algorithm by means of front-propagation in a thyristor-based hybrid system

Abstract

Propagating fronts are generic structures in a bistable diffusion-driven system and can be used to realize neural algorithms, as e.g., the Kohonen or the neural-gas algorithm. We present an analog–digital hybrid system based on a thyristor-like structure with several gate terminals. This structure represents the continuous part in which a propagating front, separating a region of high current density from a region of low current density, is used to control the learning process of the neural algorithm. With a system containing five neurons and five gates in a quasi one-dimensional arrangement it is demonstrated that an efficient parallel operating learning process can be realized by using the winner-take-all principle and the front propagation, i.e. exploiting the intrinsic dynamics of the semiconductor device. Finally, numerical and analytical investigations of the dependency of the front velocity and its width on the load current have been performed since these are essential parameters for improving the network performance.