Feedback-tolerant method and device producing weight-adjustment factors for pre-synaptic neurons in artificial neural networks
First Claim
1. In an artificial neural network comprising a plurality of functionally connected neurons, including pre-synaptic neurons and other neurons, a method of producing or generating weight-adjustment factors for said pre-synaptic neurons, comprising:
- (a) providing a plurality of interconnected artificial neuron means for use as processing unit means, each interconnected artificial neuron means comprising;
(1) a plurality of connection means that accept respective input signals from others of said plurality of interconnected artificial neuron means or input signals from external sources,(2) means for storing or supplying respective weight values for each of said plurality of connection means,(3) modulating means for modulating the values of said input signals accepted at said plurality of connection means by said respective weight values, thereby to provide a respective modulation result for each of said plurality of connection means, and(4) an output means for storing or supplying an output signal value derived by summing the results of said modulating means in said plurality of connection means and subjecting the result to a transfer function,(b) calculating a plurality of respective influence values by modulating said respective input signal values by said respective weight values within at least one of said plurality of connection means, and(c) forming an error value for each of said pre-synaptic neurons by accumulating and storing a sum of the respective influence values calculated in said plurality of connection means to which the output signal value of each pre-synaptic neuron connects,whereby said error value will be used as a factor to adjust each of said respective weight values for said plurality of connection means, so that weight values representing connection strengths from pre-synaptic neurons and external sources will tend to be adjusted most when said interconnected artificial neuron means has exercised the greatest influence over the individual modulation results of artificial neuron means to which its output means is connected.
0 Assignments
0 Petitions
Accused Products
Abstract
In an artificial neural network a method and neuron device that produce weight-adjustment factors, also called error values (116), for pre-synaptic neurons (302a . . . 302c) that are used to adjust the values of connection weights (106 . . . 106n) in neurons (100) used in artificial neural networks (ANNs). The amount of influence a pre-synaptic neuron has had over a post-synaptic neuron is calculated during signal propagation in the post-synaptic neuron (422a . . . 422n) and accumulated for the pre-synaptic neuron (426) for each post-synaptic neuron to which the pre-synaptic neuron'"'"'s output is connected (428). Influence values calculated for use by pre-synaptic neurons may further be modified by the post-synaptic neuron'"'"'s output value (102) (option 424), and its error value (116) (option 1110).
97 Citations
18 Claims
-
1. In an artificial neural network comprising a plurality of functionally connected neurons, including pre-synaptic neurons and other neurons, a method of producing or generating weight-adjustment factors for said pre-synaptic neurons, comprising:
-
(a) providing a plurality of interconnected artificial neuron means for use as processing unit means, each interconnected artificial neuron means comprising; (1) a plurality of connection means that accept respective input signals from others of said plurality of interconnected artificial neuron means or input signals from external sources, (2) means for storing or supplying respective weight values for each of said plurality of connection means, (3) modulating means for modulating the values of said input signals accepted at said plurality of connection means by said respective weight values, thereby to provide a respective modulation result for each of said plurality of connection means, and (4) an output means for storing or supplying an output signal value derived by summing the results of said modulating means in said plurality of connection means and subjecting the result to a transfer function, (b) calculating a plurality of respective influence values by modulating said respective input signal values by said respective weight values within at least one of said plurality of connection means, and (c) forming an error value for each of said pre-synaptic neurons by accumulating and storing a sum of the respective influence values calculated in said plurality of connection means to which the output signal value of each pre-synaptic neuron connects, whereby said error value will be used as a factor to adjust each of said respective weight values for said plurality of connection means, so that weight values representing connection strengths from pre-synaptic neurons and external sources will tend to be adjusted most when said interconnected artificial neuron means has exercised the greatest influence over the individual modulation results of artificial neuron means to which its output means is connected. - View Dependent Claims (2, 3, 4, 5, 6)
-
-
7. A neuron device for use in artificial neural networks, each neuron device comprising:
-
(a.) at least one synapse comprising; (1.) an input arranged to accept an input signal value from a neuron device or from an external source, (2.) a weight memory for storing or supplying at least one synapse weight or weight value, (3.) a synapse modulator for modulating said input signal value by said weight value stored or supplied at said weight memory, thereby to produce a synapse modulation result determined by modulating said input signal by the value of said synapse weight, (4.) an influence value calculator for calculating a resultant influence value by modulating said input signal value by said weight value stored or supplied at said weight memory, and (b.) a summing unit for accumulating results of said synapse modulator in said one synapse, or plural synapses into an internal sum, (c.) an output for subjecting said internal sum to a predetermined transfer function, thereby to provide an output value, and (d.) an error-value accumulator for producing an error value by accumulating and storing influence values calculated by the influence value calculators within any synapses of post-synaptic neurons to which said output of said neuron device connects, whereby said error value will be used as a factor to adjust each respective weight value for each of said one or more synapses, so that weight values representing connection strengths from pre-synaptic neurons and external sources will tend to be adjusted most when said neuron device has exercised the greatest influence over synapse modulation results of the post-synaptic neurons to which its output is connected. - View Dependent Claims (8, 9, 10, 11, 12)
-
-
13. A neuron unit, comprising:
-
(a.) at least one synapse means, comprising; (1.) connection means for accepting an input signal value from another neuron unit or from external sources, (2.) weight means for storing or supplying at least one synapse weight or weight value, (3.) synapse modulating means for modulating said input signal value by said synapse weight, thereby to produce a synapse modulation result or weighted-input, and (4.) an influence value calculating means for calculating a resultant influence value by modulating said input signal value by the value of said synapse weight, thereby to produce an influence value, (b.) a summing means for accumulating the results from said synapse modulating means in any of said one synapse means, or plural synapses means into an internal sum, (c.) an output means for subjecting said internal sum accumulated by said summing means to a predetermined transfer function, thereby to provide an output value, and (d.) an error value means for producing an error value by summing the influence values produced by said influence value calculating means in any synapse means to which said output means is connected, whereby said error value produced by said error value means will be used as a factor to adjust each respective weight value in each of said one or more synapse means, or plural synapse means so that weight values representing connection strengths from pre-synaptic neurons and external sources will tend to be adjusted most when said neuron unit has exercised the greatest influence over the individual synapse modulation results of post-synaptic neurons to which its output means is connected. - View Dependent Claims (14, 15, 16, 17, 18)
-
Specification