Hardware implementation of quantized connection nonmonotonic neural networks and a threshold learning algorithm

摘要

To realize high information processing ability of neural networks, implementing an integrated hardware with high speed is essential. To meet these requirements, we adopt two methods. One is weight quantization and another is introduction of nonmonotonic neurons. We consider a network with 3-value {-1, 0, +1} weights, which has high capability of integration, but degrade the learning performance. To compensate the degradation, nonmonotonic neurons, which exhibit high learning ability, are introduced. However, the learning performance depends on a threshold which is one of the nonmonotonic neuron's parameter, and the optimum one depends on such as problems and network structures. Therefore, we propose a threshold learning algorithm and confirm the usefulness of the learning algorithm by numerical simulations. Moreover, we have implemented such quantized connection nonmonotonic neural networks with 20 neurons and 400 synapses including the learning module using analog circuits.