Method for using a feed forward neural network to perform classification with highly biased data

• US 5,359,699 A
• Filed: 12/02/1991
• Issued: 10/25/1994
• Est. Priority Date: 12/02/1991
• Status: Expired due to Fees

• Current Assignee: General Electric Company
• Original Assignee: General Electric Company
• Inventors: Delano, Paul A., Tong, David W.
• Primary Examiner(s): Downs, Robert W.
• Application Number:
  US07/801,328
• Time in Patent Office: 1,058 Days
• Field of Search: 395/23, 395/21, 395/22, 382/14
• US Class Current: 706/20
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

An artificial neural network detects points in feature space outside of a boundary determined by a set of sample data. The network is trained using pseudo data which compensates for the lack of original data representing "abnormal" or novel combinations of features. The training process is done iteratively using a net bias parameter to close the boundary around the sample data. When the neural net stabilizes, the training process is complete. Pseudo data is chosen using several disclosed methods.

Claims

1. A method for using an artificial neural network comprising a plurality of weighted interconnected processing elements forming an input layer, an output layer, and a hidden layer connecting the input layer and the output layer, for performing classification of sensor-based data provided to said network wherein the classes to be used include and inside class consisting of one or more classes for which representative sensor-based data is available and an outside class representing an abnormal or novel class for which representative sensor-based data is unavailable or scarce, said method comprising the steps of:

gathering inside data representative of said inside class;

generating pseudo data representative of said outside class;

inputting said inside data and said pseudo data to said artificial neural network;

storing said inputted inside data and said pseudo data in said artificial neural network;

training said artificial neural network to reduce the level of classification error output using said inside data and said pseudo data, said training comprising the step of setting a bias parameter for biasing an output term for identifying said inside data;

repeating said generating, inputting, storing, and training steps until an acceptable class boundary is formed around said inside data by said artificial neural network; and

classifying the sensor-based data in accordance with the accepted class boundary.

2. The method of claim 1 wherein said biasing is adjusted during said training as a function of the accuracy of said neural network in classifying said inside data correctly.

3. The method of claim 1 wherein said pseudo data is generated randomly within a bounding hypercube large enough to contain all of said inside data.

4. The method of claim 3 wherein said pseudo data includes pseudo data from previous training iterations, wherein said previously generated pseudo data is close to the boundary as determined by said neural network.

5. The method of claim 1 wherein said pseudo data is initially generated randomly on a bounding hypercube surface, said hypercube bounding said inside data and wherein said pseudo data is thereafter generated after each said repeating step within said bounding hypercube and selected from those of said pseudo data points which are not classified as inside data by said neural network.

6. The method of claim 1 wherein said pseudo data is generated on fixed regular matrix points within a hypercube bounding said inside data.

7. The method of claim 1 wherein said training is performed using backpropagation of errors.

8. The method of claim 1 wherein said training is performed using a gradient descent method.

9. The method of claim 1 wherein said training is performed using a conjugate gradient method.

10. A method for using an artificial neural network comprising a plurality of weighted interconnected processing elements forming an input layer, an output layer, and a hidden layer connecting the input layer and the output layer, for performing classification of sensor-based data provided to said network wherein the classes to be used include a plurality of inside classes for which representative sensor-based data is available and an outside class representing an abnormal or noel class for which representative sensor-based data is unavailable or scarce, said method comprising the steps of:

gathering inside data representative of said plurality of inside classes;

generating pseudo data representative of said outside class;

inputting said inside data and said pseudo data to said artificial neural network;

storing said inputted inside data and said pseudo data in said artificial neural network;

training said neural network to reduce the level of classification error output using said inside data and said pseudo data, said training comprising the step of setting a bias parameter for biasing an output term for identifying said inside data;

repeating said generating, inputting, storing, and training steps until acceptable respective class boundaries are formed around said inside data representative of said plurality of inside classes by said artificial neural network; and

classifying the sensor-based data in accordance with the accepted class boundary.