The authors describe the design and test of an artificial neuralnnetwork, using a pulse-stream approach, that is implemented using BiCMOSntechnology. Networks are constructed from arrays of customised neuronnchips and synapse chips. The neuron chip uses novel circuitry tonimplement an accurate sigmoid transfer characteristic. The synapse chipnuses a new pulse-stream implementation of the differential amplifier andnrequires only five transistors to produce a linear multiplier. Measurednresults from the chips show that the neuron has an accurate sigmoidntransfer characteristic and gradient suitable for the errornbackpropagation learning algorithm. The synapse has excellent 1%nlinearity and properties suitable for multiplication. The chips havenbeen used to implement a three-layer artificial neural network which hasnbeen tested using hard learning problems
展开▼