Electronic ‘Neural’ Nets for Solving Ill-Posed Problems with an Entropy Regulariser

Abstract

Multiply connected analogue electronic circuits (‘neural’ nets) are characterised by having a large number of simple processing nodes such as threshold or summing devices which are connected to many other nodes through weighted interconnects. Under certain conditions the transient and equilibrium behaviour of these nets can be described in terms of a stability or Lyapunov function which is minimised at the equilibrium conditions of the net. This paper describes a circuit which is capable of solving ill-posed problems through use of an informational entropy regulariser which is incorporated into the stability function of the net. A circuit has been constructed which provides maximum entropy solutions to the loaded dice problem. The performance in terms of accuracy and speed of such circuits is discussed.