A low-cost application-specific neural network implementation with floating gate weights

The technology and design philosophy for low-cost application-specific neural network implementations through MOSIS are presented. The advantage of applying a low-cost custom-made special purpose network for a given problem over the application of available general purpose networks is discussed. The special purpose networks use floating gate devices for weight storage. The design approach presented allows the designer to implement any knowledge about the problem to be solved into hardware, thus reducing circuit complexity and training time. An example of a simple network for linear and nonlinear voltage-to-current conversion illustrate the technique