Method and apparatus for classifying patterns by use of neural network

• US 5,577,166 A
• Filed: 07/20/1992
• Issued: 11/19/1996
• Est. Priority Date: 07/25/1991
• Status: Expired due to Term

• Current Assignee: Hitachi, Ltd.
• Original Assignee: Hitachi, Ltd.
• Inventors: Mizuno, Hirotaka
• Primary Examiner(s): Knepper, David D.
• Assistant Examiner(s): Dorvil, Richemond
• Application Number:
  US07/916,541
• Time in Patent Office: 1,583 Days
• Field of Search: 395/25, 395/22, 395/21, 395/23, 395/24, 364/401, 364/402, 382/14-36
• US Class Current: 706/20
• CPC Class Codes: G06F 18/24 Classification techniques
• CPC Class Codes: G06F 18/41 Interactive pattern learning with a human teacher
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: Y10S 706/925 Business

Abstract

A method of and an apparatus for classifying an input pattern as a classification object by use of a neural network and transforming the input pattern into an output pattern as a result of classification, wherein there is obtained from the neural network having received an input pattern an output pattern for the input pattern such that the obtained output pattern is compared with a correct output pattern thereafter attained for the input pattern, a result of the comparison is sequentially stored in a memory, and the stored result of comparison is monitored to thereby detect an abnormality thereof.

Claims

1. A pattern classification method using a neural network, the method comprising steps of:

I) training the neural network including;

a) applying known training data to the neural network, the known training data each including a training input pattern which is applied to the neural network to produce a network output pattern and a corresponding reference output pattern;

b) adjusting the neural network until the network output pattern matches the reference output pattern;

c) repeating steps (a)-(b) for a plurality of training data to train the neural network;

II) classifying input patterns with the trained neural network including;

d) applying an input pattern to be classified to the trained neural network;

e) obtaining a corresponding network output pattern corresponding to the input pattern;

f) storing the input pattern and the corresponding network output pattern in a memory;

III) determining a trained neural network abnormality including;

g) comparing the stored network output pattern with a correct output pattern when the correct output pattern is subsequently obtained to determine a deviation therebetween;

h) repeating steps (d)-(g) for a plurality of input patterns to determine a plurality of the deviations;

i) storing a cumulative deviation in the memory; and

j) monitoring the cumulative deviation to detect a cumulative abnormality of the trained neural network.

2. A pattern classification method using a neural network, the method comprising steps of:

I) training the neural network with training data that includes a plurality of pairs of preselected corresponding training input and output patterns, the training including;

a) applying one of the training input patterns to the neural network;

b) adjusting the neural network such that the applied training input pattern produces the corresponding training output pattern;

c) repeating steps (a)-(b) for each of training data, whereby the neural network is trained;

II) classifying input patterns including;

d) applying an input pattern to be classified to the trained neural network and obtaining a corresponding projected output pattern;

III) determining whether to retrain the neural network including;

e) comparing the projected output pattern with a correct output pattern corresponding to the input pattern to determine a deviation therebetween;

f) repeating steps (c)-(e) for a plurality of input patterns to be classified;

g) combining the deviations to determine a cumulative deviation;

h) in response to the cumulative deviation exceeding a preselected deviation level, notifying an operation to re-train the neural network.

3. A method according to claim 2, wherein the notifying step includes:

generating a human readable notice that the neural network is due for retraining on a display.

4. A method according to claim 2, wherein:

in the comparing step the deviation is an error between the projected output pattern and the correct output pattern; and

the combining step comprises;

averaging the errors to obtain an average error; and

the comparing step includes determining whether the average error exceeds a predetermined threshold value.

5. A method according to claim 2, further comprising the step of:

generating new training data in response to the detection of the cumulative deviation exceeding the preselected level and storing the new training data.

6. A method according to claim 5, wherein the step of generating new training data generates the new training data based on previous training data used in a previous training and a previous input pattern and a previous correct output pattern obtained after the previous training.

7. A method according to claim 2, further comprising the step of re-training the neural network in response to the notifying.

8. A pattern classification apparatus using a neural network, the apparatus comprising:

means for applying each of a plurality of preselected training input patterns to the neural network, each of the preselected training input patterns having a corresponding reference output pattern;

a neural network training means for adjusting the neural network such that each preselected training input pattern produces the corresponding reference output pattern;

means for applying an input pattern to be classified to the neural network each input pattern producing a neural output pattern;

means for retrieving from a data file, a correct output pattern corresponding to each of the input patterns to be classified;

means for comparing each neural output pattern with the corresponding correct output pattern to determine a deviation therebetween;

memory means for accumulating and storing the deviations; and

means for monitoring the stored deviations to detect an abnormality condition.

9. A pattern classification apparatus using a neural network, the pattern classification apparatus comprising:

means for applying each of a plurality of known training input patterns to the neural network;

means for adjusting the neural network until the neural network produces a preselected reference output pattern corresponding to each training input pattern;

a classifying means for applying an input pattern to be classified to the neural network and obtaining a projected output pattern corresponding to the input pattern;

an abnormality condition determining means including;

means for subsequently retrieving a second output pattern from a data file, the second output pattern representing a correct output pattern corresponding to the input pattern applied for classification;

means for comparing the projected output pattern with the second output pattern corresponding to each input pattern applied for classification to determine a deviation therebetween;

memory for cumulatively storing the deviation;

means for monitoring the stored cumulative deviation to detect the abnormality condition; and

means for notifying to re-train the neural network in response to a detection of the abnormality condition.

10. An apparatus according to claim 9, wherein the notifying means includes a means for generating a human-readable display.

11. A apparatus according to claim 9, wherein the deviation represents an error between the projected output pattern and the second output pattern and the abnormality detecting means further comprises:

means for averaging the deviations to generate an average deviation representing an average of errors attained for a plurality of the input patterns applied for classification; and

means for determining whether the average deviation exceeds a predetermined threshold value.

12. An apparatus according to claim 9, further comprising:

means for generating new training data in response to the detection of the abnormality condition; and

a means for storing the new training data.

13. An apparatus according to claim 12, wherein the means of generating new training data includes:

means for generating the new training data based on previous training data used in a previous training and a previous input pattern and a previous correct output pattern obtained after the previous training.

14. An apparatus according to claim 9, further comprising:

means for re-training of the neural network in response to the means for notifying.