Adaptation method for data processing system
AssigneeYozan Inc.
InventorsSunao Takatori, Ryohei Kumagai, Koji Matsumoto, Makoto Yamamoto
Life of this patent
Estimating the term…
Abstract
An adaptation method for a data processing system comprising a plurality of neurons each of which outputs an output according to a comparison between a sum of multiplied inputs by weights and a threshold, characterized in the threshold of the neuron which has generated significant output at certain point of time is compulsorily increased to a maximal value and the weight of the neuron is adapted for a constant value of the inputs wherein the threshold is decreased to a value at the point.
Claims21
A method for maximizing an efficiency of a data processing system receiving training data, said method comprising the steps of:
- inputting said training data to said system, said system comprising a neural network having a plurality of neurons each of which is capable of outputting a signal when a weighted sum of inputs has a predetermined relationship to a threshold;
- determining a first neuron which outputs a signal in response to said training data, and compulsorily increasing a threshold of said first neuron up to a maximum value;
- determining a second neuron which did not output a signal in response to said training data, and adapting weights of said second neuron so that weights to be multiplied to inputs of the second neuron are increased when said threshold of said first neuron is increased;
- and
- decreasing said threshold of said first neuron after adapting said weights of said second neuron, thereby adapting said neural network to provide an even ignition pattern without generating a local minimum.
A method for maximizing an efficiency of a data processing system receiving training data, said method comprising the steps of:
- inputting said training data to said system, said system comprising a neural network having a plurality of neurons each of which is capable of outputting a signal when a weighted sum of inputs has a predetermined relationship to a threshold;
- determining a first neuron which outputs a signal in response to said training data, and decreasing a weight of said first neuron, said weight being compulsorily decreased to a minimum value;
- determining a second neuron which did not output a signal in response to said training data, and adapting weights of said second neuron so that weights to be multiplied to inputs of the second neuron are decreased when said weight of said first neuron is decreased;
- and
- increasing said weight of said first neuron after adapting said weights of said second neuron, thereby adapting said neural network to provide an even ignition pattern without generating a local minimum.
A method for converting an untrained neural network into a trained neural network comprising the steps of:
- inputting a first training signal into a data processing system having first and second neurons, said first neuron receiving the first training signal and said second neuron receiving an output from the first neuron, each neuron being capable of generating an output signal when a weighted sum of inputs has a predetermined relationship to a threshold;
- increasing a threshold of the first neuron to a compulsory level when the first neuron ignites;
- inputting a second training signal into the data processing system;
- adapting a weight of the second neuron in response to the second training signal and in response to the output produced by the first neuron at a time when the threshold of the first neuron is at the compulsory level;
- decreasing the threshold of the first neuron, thereby generating a trained neural network in the data processing system.
Dependencies are inferred from the claim text.