In this dissertation, a modular analog neural network chip set with on-chip learning capability is presented. These synapse chip and neuron chip are designed with modular concepts and expandable for multi-layer and large numbers of neurons and synapses. The expansion is so easy that it does not require any additional circuits and is completed by pin-to-pin connections. As a learning algorithm, error back-propagation is incorporated in the chip set. For given input and output patterns, the neural network chip adapts by itself. The chips are fabricated using 0.8㎛ CMOS process and are used to implement analog neural network system. The system is proven to have ability to learn gray patterns and character recognition as well as XOR problems.
These analog neural network system is applied to a active noise control. Basic idea of the active noise control is fully investigated with simulations. From the simulation results, the active noise control using neural network shows better performance and stability than that using transversal adaptive filter, which gives a good reason to implement and apply neural network chip to the active noise control. The analog neural network system is successfully trained to cancel noise in feedforward active noise control.
Adaptive equalizing which does important roles in mobile communications is also one of good applications. Simulations for equalizer are performed to verify superior of the neural network to conventional schemes in minimum and nonminimum phase channels. The adaptive equalizer using the analog neural network system is adaptive is proven to diminish ISI (intersymbol interference) caused by dispersive channel and shows possibility to be used in communication receivers.