Pattern recognition apparatus and pattern learning apparatus employing neural net including excitatory element-inhibitory element pair couplings

• US 5,323,471 A
• Filed: 03/30/1992
• Issued: 06/21/1994
• Est. Priority Date: 09/12/1991
• Status: Expired due to Fees

• Current Assignee: ATR Auditory and Visual Perception Research Laboratories
• Original Assignee: ATR Auditory and Visual Perception Research Laboratories
• Inventors: Hayashi, Yukio
• Primary Examiner(s): Boudreau, Leo H.
• Application Number:
  US07/859,795
• Time in Patent Office: 813 Days
• Field of Search: 382/14, 382/15, 395/21, 395/24, 395/23
• US Class Current: 382/158
• CPC Class Codes: G06V 30/194 References adjustable by an adaptive method, e.g. learning

Abstract

Disclosed in a pattern recognition apparatus including a neural net including a plurality of mutually coupled excitatory element-inhibitory element paris. The elements in the neural net are coupled through pair coupling coefficients and an excitatory element coupling coefficient. A teacher signal generator outputs a successive teacher signal Q (t) and its infinitesimal variation ΔQ (t) in accordance with a learning pattern, to apply the output teacher signal and the output small variation to a coupling coefficient controller. The coupling coefficient controller carries out a learning processing in accordance with the teacher signal, so that the coupling coefficients in the neural net are updated. An operation time necessary for a pattern recognition is shortened. In addition, since the learning processing is carried out by only a forward processing on time base, a learning time is also decrease.

Claims

1. A pattern recognition apparatus for recognizing an applied input pattern, said input pattern including a plurality of input signals for defining said input pattern, said apparatus comprising:

a neural net including a plurality of excitatory element-inhibitory element pairs which are mutually coupled with each other in accordance with a predetermined relation,

each said element pair having a corresponding excitatory element and a corresponding inhibitory element coupled in opposite directions through a first pair coupling coefficient for the excitatory coupling and a second pair coupling coefficient for the inhibitory coupling, respectively,

said plurality of excitatory elements each receiving said plurality of input signals,

any two of said plurality of excitatory elements being coupled in opposite directions through excitatory element coupling coefficients;

teacher signal generating means for generating a teacher signal in accordance with a learning pattern to be learned;

coupling coefficient updating means responsive to a teacher signal for updating said first and second pair coupling coefficients and said excitatory element coupling coefficients in said neural net,

after updating by said coupling coefficient updating means, said neural net receiving said plurality of input signals and outputting an activation signal through said plurality of excitatory elements; and

classification determining means responsive to said activation signal for determining classification of said input pattern.

2. The pattern recognition apparatus according to claim 1, wherein

said teacher signal generating means includes learning pattern changing means for generating said teacher signal by continuously changing said learning pattern on the basis of a predetermined rule.

3. The pattern recognition apparatus according to claim 2, wherein

said learning pattern includes a learning image pattern placed on a two-dimensional virtual image, and

said learning pattern changing means includes,

image skewing means for skewing said virtual image in a predetermined ratio, and

image distortion means for distorting said virtual image based on a predetermined expression, wherein

said learning image pattern is varied in accordance with said virtual image changed by said image skewing means and said image distortion means.

4. The pattern recognition apparatus according to claim 3, wherein

said learning pattern is varied discretely, and

said learning pattern changing means further includes continuous conversion means for converting said learning image pattern into a continuously time-sequential signal.

5. The pattern recognition apparatus according to claim 1, wherein

said classification determining means includes periodicity determining means for determining whether the activation signal output from said neural net has periodicity or not, and

the classification of said input pattern is determined in response to said periodicity determining means.

6. The pattern recognition apparatus according to claim 5, wherein

said classification determining means further includes correlation distance detecting means for detecting a correlation distance between said activation signal and a predetermined dictionary pattern when said periodicity determining means determines the periodicity of the activation signal, and

the classification of said input pattern is determined in response to said correlation distance detecting means.

7. The pattern recognition apparatus according to claim 1, wherein

said neural net further includes,

a plurality of first multipliers for multiplying an output signal of corresponding one of said excitatory elements by corresponding one of said first pair coupling coefficients and applying the multiplied signal to corresponding one of said inhibitory elements,

a plurality of second multipliers for multiplying an output signal of corresponding one of said inhibitory elements by corresponding one of said second pair coupling coefficients and applying the multiplied signal to corresponding one of said excitatory elements, and

a plurality of third multipliers for multiplying respective output signals of any two of said plurality of excitatory elements by corresponding ones of said excitatory element coupling coefficients, respectively, and applying one multiplied signal and the other multiplied signal to the other and one of said two output signals, respectively.

8. The pattern recognition apparatus according to claim 3, wherein

said learning image pattern comprises a plurality of learning character patterns each indicating different characters, and

said classification determining means includes character selecting means responsive to said activation signal for selecting the classification of said input pattern from said plurality of learning character patterns.

9. The pattern recognition apparatus according to claim 1, wherein

said plurality of input signals for defining said input pattern comprise a plurality of mesh density feature signals for defining a provided two-dimensional image.

10. The pattern recognition apparatus according to claim 7, wherein

said neural net further includes a plurality of fourth multipliers for multiplying corresponding one of said input signals by a corresponding input coupling coefficient and applying the multiplied signal to corresponding one of said excitatory elements, and

said coupling coefficient updating means further updates the input coupling coefficient in said neural net.

11. The pattern recognition apparatus according to claim 1, wherein

said coupling coefficient updating means updates the second pair coupling coefficient and the excitatory element coupling coefficient in said neural net by carrying out only a forward processing on time base.

13. The pattern recognition apparatus according to claim 1, wherein said plurality of input signals for defining said input pattern comprise a plurality of direction contributivity density feature signals for defining a provided two-dimension image.

12. A pattern learning apparatus, comprising:

a neural net including a plurality of excitatory element-inhibitory element pairs which are mutually coupled with each other in accordance with a predetermined relation,

each said element pair having corresponding one of said excitatory elements and corresponding one of said inhibitory elements coupled with each other in opposite directions through first and second pair coupling coefficients,

any two of said plurality of excitatory elements being coupled in opposite directions through an excitatory element coupling coefficient;

teacher signal generating means for generating a teacher signal in accordance with a learning pattern to be learned; and

sequential-time learning processing means for updating said second pair coupling coefficient and said excitatory element coupling coefficient in said neural net by carrying out only a forward processing on a time base with respect to the teacher signal.

14. A pattern recognition apparatus for recognizing an applied input pattern, said input pattern including a plurality of input signals for defining said input pattern, said apparatus comprising:

a neural net including a plurality of excitatory element-inhibitory elements pairs which are mutually coupled through with each other in accordance with a predetermined relation,

each of said element pair having a corresponding excitatory element and a corresponding inhibitory element coupled in opposite directions through a first pair coupling coefficient for the excitatory coupling and a second pair coupling coefficient for the inhibitory coupling, respectively,

said plurality of excitatory elements each receiving said plurality of input signals,

any two of said plurality of excitatory elements being coupled in opposite directions through excitatory element coupling coefficients;

said neural net receiving said plurality of input signals and outputting an activation signal through said plurality of excitatory elements; and

classification determining means responsive to said activation signal for determining classification of said input pattern.