Pattern recognition neural net insensitive to disturbances in inputs

• US 5,396,565 A
• Filed: 11/04/1992
• Issued: 03/07/1995
• Est. Priority Date: 06/04/1990
• Status: Expired due to Fees

• Current Assignee: NEC Corporation
• Original Assignee: NEC Corporation
• Inventors: Asogawa, Minoru
• Primary Examiner(s): Couso, Yon J.
• Application Number:
  US07/971,253
• Time in Patent Office: 853 Days
• Field of Search: 382/14, 382/15, 364/513, 364/807, 395/20, 395/21, 395/22, 395/23, 395/24
• US Class Current: 382/159
• CPC Class Codes: G06V 30/194 References adjustable by an adaptive method, e.g. learning

Abstract

In a neural network assembly for use in pattern recognition, a memory region memorizes reference patterns along with their categories. A pattern associator neural network is connected to the memory region and is trained in accordance with a back-propagation training algorithm in a training phase of operation to correctly recognize the reference patterns according to the categories. An adaptive input region is connected to the neural network, supplied with roughly segmented patterns as input patterns, and trained in a recognition phase of operation following the training phase. and correctly processes the input patterns into processed patterns which the neural network can correctly recognize according to the categories. Preferably, the assembly is operable in the training phase as comprising the memory region, the neural network, and a controlling part which is connected to the memory region and the neural network to train the neural network. In the recognition phase the assembly is operable as comprising the input region, the neural network and the controlling part which is now connected to the input region and the neural network and adjusts the input region so that each roughly segmented pattern may approach a pertinent one of the reference patterns.

Claims

1. A neural network assembly, comprising:

a memory region for storing reference patterns and categories of said reference patterns;

an adaptive input region, having region cells, for adaptively processing input patterns having disturbances into processed patterns with said disturbances removed, said adaptive input region being trainable to remove said disturbances by adjustment of said region cells of said adaptive input region by resilient interconnection;

a pattern associator neural network, selectively coupled to said memory region and said adaptive input region, for recognizing patterns in said processed patterns corresponding to said reference patterns and for producing a recognition result representative of at least one of said categories;

.

a training and recognition controlling unit, coupled to said memory region and said pattern associator neural network, for training said pattern associator neural network to correctly produce said recognition result and coupled to said adaptive input region for training said adaptive input region to remove said disturbances by said adjustment of said region cells by resilient interconnection;

wherein said training and recognition controlling unit provides a training phase of operation and a recognition phase of operation of said pattern associator neural network, said recognition phase following said training phase; and

further comprising a switching network for coupling said pattern associator neural network to said memory region during said training phase and for coupling said pattern associator neural network to said adaptive input region during said recognition phase.

2. A neural network assembly, comprising:

a memory region for storing reference patterns and categories of said reference patterns;

an adaptive input region, having region cells, for adaptively processing input patterns having disturbances into processed patterns with said disturbances removed, said adaptive input region being trainable to remove said disturbances by adjustment of said region cells of said adaptive input region by resilient interconnection;

a pattern associator neural network, selectively coupled to said memory region and said adaptive input region, for recognizing patterns in said processed patterns corresponding to said reference patterns and for producing a recognition result representative of at least one of said categories;

a training and recognition controlling unit, coupled to said memory region and said pattern associator neural network, for training said pattern associator neural network to correctly produce said recognition result and coupled to said adaptive input region for training said adaptive input region to remove said disturbances by said adjustment of said region cells by resilient interconnection; and

a switching network, controlled by said controlling unit, for coupling said memory region to said pattern associator neural network for enabling said controlling unit during a training phase to train said pattern associator neural network by using said reference patterns and said categories from said memory region, and for coupling said adaptive input region to said pattern associator neural network for enabling said controlling unit during a recognition phase to train said adaptive input region.

3. A neural network assembly comprising:

a memory region for memorizing reference patterns and categories of said reference patterns;

an adaptive input region for adaptively processing input patterns having disturbances into processed patterns with said disturbances removed, said adaptive input region having a plurality of region cells bound together with resilient connections in accordance with an address force and being trainable to remove said disturbances by adjustment of said resilient connections;

a pattern associator neural network, selectively coupled to said memory region and said adaptive input region, for recognizing patterns in said processed patterns corresponding to said reference patterns and for producing a recognition result representative of at least one of said categories;

a training and recognition controlling unit, coupled to said memory region and said pattern associator neural network, for training said pattern associator neural network to correctly produce said recognition result and coupled to said adaptive input region for providing a delta force input for training said adaptive input region to remove said disturbances by said adjustment of said resilient connections of said region cells; and

a switching network, controlled by said controlling unit, for coupling said memory region to said pattern associator neural network for enabling said controlling unit during a training phase of operation to train said pattern associator neural network by using said reference patterns and said categories from said memory region, and for coupling said adaptive input region to said pattern associator neural network for enabling said controlling unit during a recognition phase of operation to train said adaptive input region.

4. A neural network assembly, as in claim 3, wherein said training and recognition controlling unit provides said training phase of operation and said recognition phase of operation of said pattern associator neural network, said recognition phase following said training phase.