A modular analog chip for feed-forward networks with on-chip learning

A feedforward artificial neural network (ANN) with learning capability in modular design is presented using CMOS circuits. We employ a modified error backpropagation continuous-time learning rule. A (nonlinear) analog Gilbert multiplier is used as a synapse and a wide-range transconductance amplifier is used as soma. For learning circuits, the, same multiplier is used for updating the weights. Test results demonstrate the successful operation of the chip. Finally, a modular chip design for a large scale implementation of feedforward ANN with learning is described.<
>