EP0384689A2 — A learning system for a data processing apparatus

Claim 1

Independent Claim 216 words

A learning system for use in a data processing apparatus for learning an input pattern by obtaining an internal-state value necessary for realizing a desired data conversion, for performing a pattern conversion defined by said internal-state value and for calculating an output pattern corresponding to said input pattern, comprising: pattern presenting means for presenting an input pattern group of the subject to be learned for said pattern conversion, dividing the input pattern group of the subject to be learned into at least two sets, selecting one of the divided sets, presenting the input patterns of the selected set to said pattern conversion and then presenting the input patterns belonging to all the sets presented up to the current point when the internal-state value to be converged is obtained in accordance with the presentation of said selected set, error value calculating means for calculating an error value representing a magnitude of a non-consistency between an output pattern outputted in accordance with said presentation and a teacher pattern representing a pattern to be obtained by said output pattern, and an internal-state value updating means for updating an internal-state value in accordance with an updated quantity of the internal-state value calculated based on the error value and obtaining an internal-state value whose error value holds within a specified allowance.

Claim 2

Dependent Claim 216 words

A learning system for use in a data processing apparatus for learning an input pattern by obtaining an internal-state value necessary for realizing a desired data conversion, for performing a pattern conversion defined by said internal-state value and for calculating an output pattern corresponding to said input pattern, comprising: pattern presenting means for presenting an input pattern group of the subject to be learned for said pattern conversion, dividing the input pattern group of the subject to be learned into at least two sets, selecting one of the divided sets, presenting the input patterns of the selected set to said pattern conversion and then presenting the input patterns belonging to all the sets presented up to the current point when the internal-state value to be converged is obtained in accordance with the presentation of said selected set, error value calculating means for calculating an error value representing a magnitude of a non-consistency between an output pattern outputted in accordance with said presentation and a teacher pattern representing a pattern to be obtained by said output pattern, and an internal-state value updating means for updating an internal-state value in accordance with an updated quantity of the internal-state value calculated based on the error value and obtaining an internal-state value whose error value holds within a specified allowance.

Claim 3

Dependent Claim 41 words

The learning system according to claim 1, wherein said internal-state value updating means uses an internal-state value obtained from the previously presented input pattern set as an initial value when the internal-state value is learnt in accordance with said pattern presentation.

Claim 4

Dependent Claim 65 words

The learning system according to claim 1, wherein when a pattern or a plurality of input patterns which are a subject of a learning are added during performance of the learning process after completion of the learning of the internal state value, said pattern presenting means presents said added input pattern as one set for said pattern conversion, thereby further learning the internal state value.

Claim 5

Dependent Claim 73 words

The learning system according to claim 1, wherein when a pattern or a plurality of input patterns which are a subject of a learning are added during performance of the learning process after completion of the learning of the internal state value, said pattern presenting means presents said added input pattern by dividing said added input patterns into a plurality of sets for said pattern conversion, thereby further learning the internal state value.

Claim 6

Dependent Claim 93 words

The learning system according to claim 1, wherein when a pattern or a plurality of input patterns which are a subject of a learning are added during performance of the learning process after completion of the learning of the internal state value, said pattern presenting means presents said added input pattern by combining said added input pattern with the original input pattern for said pattern conversion so that said added input pattern and the original input pattern are divided into a plurality of input pattern sets, thereby further learning the internal state value.

Claim 7

Dependent Claim 152 words

The learning system according to claim 1, wherein said pattern conversion is conducted by a basic unit for receiving one or a plurality of inputs from the previous stage layer and an internal state value to be multiplied by said input, thereby providing the sum of the products, and for converting the sum of the products thus obtained to a final output through a threshold value function, an input layer comprises a plurality of sets of basic units, one or a plurality of stages of hidden layers, each comprising a plurality of basic units, an output layer comprises one or a plurality of said basic units, an internal connection is formed between the input layer and the first hidden layer, between hidden layers, and between the last hidden layer and the output layer, and the hidden state value is set in accordance with the internal connection, thereby forming a layered network structure.

Claim 8

Dependent Claim 224 words

A learning system for learning an input pattern by obtaining an internal state value for realizing a data conversion and by calculating an output pattern corresponding to an input pattern by performing a data conversion defined by an internal state value, comprising: pattern presentation means for presenting an input pattern group of a learning subject for said pattern conversion, error value calculating means for calculating an error representing an amplitude of non-consistency between an output pattern group outputted in accordance with said presentation and a teacher pattern group presenting a pattern to be determined as the output pattern group, internal-state value updating means for obtaining an internal state value whose error value holds within a specified range by updating the internal state value in accordance with the updated quantity of the internal state value calculated based on the error value, convergence judgement means for judging whether respective output patterns corresponding to the input pattern group presented by said pattern presentation means in accordance with progress in the updating process of the internal state value converge to the corresponding teacher patterns, and convergence pattern excluding means for excluding a value relating to the output pattern subjected to said convergence from a calculation for updating the internal state value when the internal state value is judged as a converged output pattern by said convergence judgement means.

Claim 9

Dependent Claim 224 words

A learning system for learning an input pattern by obtaining an internal state value for realizing a data conversion and by calculating an output pattern corresponding to an input pattern by performing a data conversion defined by an internal state value, comprising: pattern presentation means for presenting an input pattern group of a learning subject for said pattern conversion, error value calculating means for calculating an error representing an amplitude of non-consistency between an output pattern group outputted in accordance with said presentation and a teacher pattern group presenting a pattern to be determined as the output pattern group, internal-state value updating means for obtaining an internal state value whose error value holds within a specified range by updating the internal state value in accordance with the updated quantity of the internal state value calculated based on the error value, convergence judgement means for judging whether respective output patterns corresponding to the input pattern group presented by said pattern presentation means in accordance with progress in the updating process of the internal state value converge to the corresponding teacher patterns, and convergence pattern excluding means for excluding a value relating to the output pattern subjected to said convergence from a calculation for updating the internal state value when the internal state value is judged as a converged output pattern by said convergence judgement means.

Claim 10

Dependent Claim 42 words

The leaning system according to claim 7, wherein said internal state value updating means calculates an updated quantity of a calculated internal state value such that the updated quantity increases as the number of output patterns which are judged as non-convergent decreases.

Claim 11

Dependent Claim 50 words

The learning system according to claim 7, wherein a convergence judgement means carries out a judgement of convergence in accordance with an intermittent cycle among cycles for updating the internal state value and said convergence pattern excluding means excludes the converged output pattern during the period of said intermittent cycle.

Claim 12

Dependent Claim 141 words

The learning system according to claim 7, wherein said pattern conversion is conducted by using a basic unit for receiving one or a plurality of inputs from said prestage layer and an internal state value is multiplied by said input, thereby providing a sum of the products, and converting said sum of the products into a threshod function, thereby providing a final output, an input layer comprises a plurality of basic units, one or a plurality of stages of an hidden layer comprises a plurality of said basic units, an internal connection is made between an input layer and an hidden layer of the first stage, between hidden layers, and between an hidden layer of the last stage and the output layer, thereby determining the internal state value in accordance with the internal connection, and thereby providing a layered network structure.

Claim 13

Dependent Claim 83 words

A learning system for use in a data processing apparatus for learning an input pattern by obtaining an internal-state value necessary for realizing a desired data conversion comprising: means for dividing an input learning pattern group into at least two sets, a first means for presenting a first set of the input learning pattern for the data conversion to obtain a converged internal state value, and a second means for presenting a second set of the input learning pattern for the data conversion.

Claim 14

Dependent Claim 83 words

A learning system for use in a data processing apparatus for learning an input pattern by obtaining an internal-state value necessary for realizing a desired data conversion comprising: means for dividing an input learning pattern group into at least two sets, a first means for presenting a first set of the input learning pattern for the data conversion to obtain a converged internal state value, and a second means for presenting a second set of the input learning pattern for the data conversion.

Claim 15

Dependent Claim 33 words

The learning system according to claim 11 further comprising; means for unifying said input pattern belonging to said first and second set of the learning pattern to present them for said data conversion.

Claim 16

Dependent Claim 32 words

The learning system according to claim 12, wherein said dividing means and said unifying means are combined in an appropriate manner to learn the input pattern belonging to said input pattern group.

Claim 17

Dependent Claim 204 words

A learning method for use in a data processing apparatus for learning an input pattern by obtaining an internal-state value necessary for realizing a desired data conversion by performing a pattern conversion defined by said internal-state value and calculating an output pattern corresponding to said input pattern, comprising the steps of presenting an input pattern group of the subject to be learned for said pattern conversion, dividing the input pattern group of the subject to be learned into at least two sets, selecting one of the divided sets, presenting the input pattern group of the selected set to said pattern conversion and sequentially presenting an input pattern group belonging to all the sets presented up to the current point when the internal-state value to be converged is obtained in accordance with the presentation, calculating an error value representing a magnitude of a non-consistency between an output pattern group outputted in accordance with said presentation and a teacher pattern group representing a pattern to be obtained by said output pattern group, and updating an internal-state value in accordance with an updated quantity of the internal-state value calculated based on the error value and obtaining an internal-state value whose error value holds within a specified allowance.

Claim 18

Dependent Claim 204 words

A learning method for use in a data processing apparatus for learning an input pattern by obtaining an internal-state value necessary for realizing a desired data conversion by performing a pattern conversion defined by said internal-state value and calculating an output pattern corresponding to said input pattern, comprising the steps of presenting an input pattern group of the subject to be learned for said pattern conversion, dividing the input pattern group of the subject to be learned into at least two sets, selecting one of the divided sets, presenting the input pattern group of the selected set to said pattern conversion and sequentially presenting an input pattern group belonging to all the sets presented up to the current point when the internal-state value to be converged is obtained in accordance with the presentation, calculating an error value representing a magnitude of a non-consistency between an output pattern group outputted in accordance with said presentation and a teacher pattern group representing a pattern to be obtained by said output pattern group, and updating an internal-state value in accordance with an updated quantity of the internal-state value calculated based on the error value and obtaining an internal-state value whose error value holds within a specified allowance.

A learning system for a data processing apparatus

Abstract

Inventors

Bibliographic Details

Claims Analysis

References Cited (Backward Citations)

USPTO Annuity & Maintenance Windows

Assigned Categories & Tags

Assign Category Tag