Parallel multi-value neural networks

• US 5,768,476 A
• Filed: 03/29/1995
• Issued: 06/16/1998
• Est. Priority Date: 08/13/1993
• Status: Expired due to Fees

• Current Assignee: KDDI Corporation
• Original Assignee: Kokusai Denshi Denwa Company Limited (KDDI Corporation)
• Inventors: Sugaya, Fumiaki, Yatsuzuka, Youtaro
• Primary Examiner(s): Hafiz, Tariq R.
• Assistant Examiner(s): Smith, Jeffrey S.
• Application Number:
  US08/406,905
• Time in Patent Office: 1,175 Days
• Field of Search: 364/737, 364/739, 364/746, 364/746.2, 395/20-23
• US Class Current: 706/26
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/065 Analogue means
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

In a parallel multi-value neural network having a main neural network 16 and a sub neural network 18 coupled with the main neural network 16 in parallel for an input signal, the main neural network 16 is trained with a training input signal by using a main multi-value teacher signal, and the sub neural network is successively trained with the training input signal by using multi-value errors between a multi-value output signal of the main neural network 16 derived through a multi-value threshold means 17 and the main multi-value teacher signal, so as to compensate the multi-value errors involved in the multi-value output signal of the main neural network 16 by the multi-value output signal of the sub neural network 18 derived through a multi-value threshold means 19. A desired multi-value output signal of the parallel multi-value neural network 15 is obtained by adding in modulo the multi-value output signals of both the neural networks through a multi-value modulo adder 20.

Claims

1. A parallel multi-value neural network comprising:

a main neural network which is trained at first with a training input signal by using a main multi-value teacher signal;

at least one sub neural network coupled with said main neural network in parallel for an input signal;

at least two multi-value threshold means, wherein a first multi-value threshold means for providing a multi-value output signal of said main neural network by quantizing an output of said main neural network into a multi-value, and a second multi-value threshold means for providing a multi-value output signal of said at least one sub neural network by quantizing an output of said at least one sub neural network into a multi-value,

said at least one sub neural network being trained with said training input signal by using a compensatory multi-value teacher signal, said compensatory multi-value teacher signal is obtained by converting at least a part of multi-value errors to a predetermined code system having codes at larger distances from each other than codes of said at least a part of multi-value errors before conversion, said multi-value errors being a difference between said main multi-value teacher signal and a multi-value output signal of said main neural network which has been trained, said multi-value output signal being derived through said first multi-value threshold means; and

at least one multi-value modulo adding means, which adds, in modulo, a) said multi-value output signal of said main neural network which has been trained, said multi-value output signal derived through said first multi-value threshold means, and b) a signal obtained by restoring said predetermined code system to an original code system in an output of said second multi-value threshold means which receives an output of said at least one sub neural network which has been trained,

said multi-value modulo adding means providing a desired multi-value output signal by compensating said multi-value errors involved in said multi-value output signal of said main neural network derived through said first multi-value threshold means.

2. A parallel multi-value neural network according to claim 1, wherein:

said predetermined code system is an error correcting code system.

3. A parallel neural network according to any of claims from 1 to 2, wherein:

a last sub neural network learns again for an increased number of hidden units when a multi-value output signal of said last sub neural network derived through said second multi-value threshold means cannot completely coincide with a compensatory multi-value teacher of said last neural network within a given number of training cycles.

4. A parallel neural network according to any of claims from 1 to 2, wherein:

either said main neural network or sub neural networks precedent to a last sub neural network learn again or learn again for increased numbers of hidden units when a multi-value output signal of said last sub neural network cannot completely coincide with a compensatory multi-value teacher of said last sub neural network within a given number of training cycles.

5. A parallel neural network according to claim 1 further comprising a mode controller means for controlling the main neural network and said at least one sub neural network to set weighing factors obtained in a learning process and to perform an execution process.