Hetero-Correlation-Associative Memory with Trigger Neurons: Accumulation of Memory through Additional Learning in Neural Networks

摘要

In this paper, we present a hetero-correlation-associative memory model that allows additional learning without destroying existing stored data, where the model is a feed-forward neural network consisting of two layers. The number of neurons in the input layer (called "trigger neurons"）increases as the number of stored images increases. One trigger neuron is linked to one image to be learned. Each time an image to be learned is added, a new trigger neuron is also added, thereby enabling the model to learn the additional image. Moreover, since the learning process simply adds new trigger neurons, it does not influence previously learned data. We can store images in the network as necessary, one after another. The stored images can be recalled through the firing of the corresponding trigger neuron. The recalled images are approximately perfect matches to the teacher images. We also show that using the proposed learning procedure greatly improves the memory rate compared with the conventional one.