Method of storing associative information and a device therefor

• US 5,586,222 A
• Filed: 10/17/1994
• Issued: 12/17/1996
• Est. Priority Date: 07/24/1991
• Status: Expired due to Fees

• Current Assignee: Mitsubishi Denki Kabushiki Kaisha (Mitsubishi Electric Corporation), Mitsubishi Electric Corporation
• Original Assignee: Mitsubishi Denki Kabushiki Kaisha (Mitsubishi Electric Corporation)
• Inventors: Zhang, Wei, Ishii, Toshinao, Takahashi, Masanobu, Kyuma, Kazuo
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Dorvil, Richemond
• Application Number:
  US08/324,378
• Time in Patent Office: 792 Days
• Field of Search: 395/4, 395/24, 395/26, 395/22, 395/27, 395/11
• US Class Current: 706/18
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks
• CPC Class Codes: G06V 10/88 Image or video recognition using optical means, e.g. reference filters, holographic masks, frequency domain filters or spatial domain filters

Abstract

An input pattern or two-dimensional pattern is associated and stored with use of associative matrices having the same size as that of an input pattern without converting it into one-dimensional pattern, wherein the associative matrices are generated from the input pattern so as to maximize the Hamming distance between state invariants which correspond to each different storage pattern, and thereby a power of categorization that corresponds to each storage pattern can be enhanced, thus increasing the storage capability and robustness.

Claims

1. An associative information storage device comprising:

an associative information storage circuit, including at least one neural network, for obtaining a state output from a state input by inputting associative matrices from a source exterior to said circuit and conversion means to obtain said state input for a succeeding time period by converting said state output using a nonlinear conversion;

initial value input means for inputting a fixed initial state to said associative information storage circuit;

storage means for storing progression of state transitions of said associative information storage circuit; and

associative matrix processing means for generating associative matrices of said neural network based on an input pattern.

2. An associative information storage device as defined in claim 1, wherein said neural network enables said associative matrices to be stored optically, wherein said associative matrices are generated optically based on said input pattern and fed to said associative information storage circuit.

3. An associative information storage device as defined in claim 1, wherein said neural network enables said associative matrices to be stored electronically, wherein said associative matrices are generated electronically based on said input pattern and fed to said associative information storage circuit.

4. An associative information storage device comprising:

an associative matrix processing unit for computing an associative matrix based on an input pattern;

neural networks for receiving said associative matrix from said associative matrix processing unit;

an initial value input unit for inputting a predetermined fixed initial input value to said neural networks as a predetermined initial state;

a state output connecting unit for receiving an output from said neural networks after a first state transition which is a product of said predetermined fixed initial input value and said associative matrix and outputting a result; and

a conversion unit for convening the result from said state output connecting unit using a nonlinear conversion function to produce a convened result.

5. A device as claimed in claim 4 wherein said associative matrix in said neural networks is changed during said first state transition.

6. A device as claimed in claim 5 further comprising means, connected to said conversion means, for inputting said converted result to said neural networks as an updated input value;

wherein said neural networks undergo a second state transition to produce a second output.

7. A method of storing associative information comprising the steps of:

computing an associative matrix of a neural network on the basis of an input pattern;

loading the computed associative matrix into said neural network as an associative matrix of said neural network;

inputting a predetermined initial state into said neural network as an initial state of said neural network;

causing said neural network to undergo a progression of state transitions starting with said predetermined initial state and said loaded associative matrix; and

storing associative information associating said input pattern and said progression of state transitions.

8. A method as claimed in claim 7 wherein said associative matrix changes during said progression of state transitions.

9. A method as claimed in claim 7 wherein said associative matrix is not set during a recognition phase of operation of said neural network.