Analog CMOS implementation of backward error propagation

Novel CMOS analog circuits for the implementation of feedforward neural networks with backward error-propagation learning are explored. Hardware learning circuitry can successfully obtain the strengths of the synaptic weights that approximately satisfy a nonlinear mapping. Weights and input values can be stored as charges on capacitors; they are periodically refreshed by interface circuits that convert values stored in digital memory into analog signals. Extensive SPICE (simulation program with IC emphasis) simulation results are presented. These circuits entail learning a set of desired input-output pairs within several hundred micro seconds. >