Learning system for a data processing apparatus

• US 6,023,693 A
• Filed: 10/06/1995
• Issued: 02/08/2000
• Est. Priority Date: 02/20/1989
• Status: Expired due to Fees

• Current Assignee: Fujitsu Limited
• Original Assignee: Fujitsu Limited
• Inventors: Masuoka, Ryusuke, Watanabe, Nobuo, Kimoto, Takashi, Tanahashi, Jun'ichi, Kawamura, Akira, Asakawa, Kazuo
• Primary Examiner(s): Downs, Robert W.
• Application Number:
  US08/540,030
• Time in Patent Office: 1,586 Days
• Field of Search: 395/23, 395/22, 395/21
• US Class Current: 706/25
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

A learning system is used 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 the internal-state value and calculating an output pattern corresponding to the input pattern. The learning system comprises a pattern presenting unit for presenting an input pattern group of the subject to be learned for 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 a pattern conversion unit and 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 of the selected set, and an error value calculating unit for calculating an error value representing a magnitude of a non-consistency between an output pattern group outputted in accordance with the presentation and a teacher pattern group representing a pattern to be obtained by the output pattern group.

Claims

1. A learning pattern presenting device in 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 pattern group to be learned into at least two sets of a plurality of divided sets; and

means for sequentially selecting a set from the plurality of divided sets, sequentially presenting each selected set to the learning system for training until convergence, and after training of the learning system for each selected sets, presenting a combination of the plurality of divided sets to the learning system for training to obtain a final internal state value.

2. A pattern learning method for use with a data processing device for converting data according to an internal state value and calculating and outputting an output pattern corresponding to an input pattern, said method comprising the steps of:

a) presenting an input pattern group to be learned;

b) calculating an error value between a teacher pattern and an output pattern group converted from the input pattern group using the internal state value;

c) determining, based on the error value, whether or not each output pattern has converged into a corresponding teacher pattern;

d) with remaining output patterns excluding the output patterns determined to have been converged as a result of said determining step c), calculating an update amount of the internal state value based on the error value such that the updated amount increases as the number of the remaining output patterns decrease; and

e) determining a next internal state value according to the calculated update amount.

3. A pattern presenting method using a learning pattern presenting device in 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, said method comprising the steps of:

a) dividing an input pattern group to be learned into at least two sets of a plurality of divided sets; and

b) sequentially selecting a set from the plurality of divided sets, sequentially presenting each selected set to the learning system for training until convergence, and after training the learning system for each selected sets presenting a combination of the plurality of divided sets to the learning system for training to obtain a final internal state value.