Tree-like perceptron and a method for parallel distributed training of such perceptrons

摘要

Constraints placed on the structure of a conventional multi- layer network consequently enable learning rules to be simplified and the probability of reaching only local minima to be reduced. These constraints include neurons which are either inhibitory or excitatory. Also, for each neuron in the hidden layer, there is at most one synapse connecting it to a corresponding neuron in the output layer. The result is a tree-like structure which facilitates implementation of large scale electronic networks, and allows for parallel training of parts of the network. Additionally, each neuron in the hidden layer receives a reinforcement signal from its corresponding neuron in the output layer which is independent of the magnitude of synapses posterior to the hidden layer neuron. There may be multiple hidden layers, wherein each layer has a plurality of neurons, and wherein each neuron in an anterior layer connects to only one neuron in any posterior layer. In training, weights of synapses connected anterior to any neuron are adjusted with the polarity opposite the polarity of the error signal when the polarity determined for the path for the neuron is inhibitory. The adjustment is made with the polarity of the error signal when the polarity determined for the path for the neuron is excitatory.