Method, apparatus and storage medium configured to analyze predictive accuracy of a trained neural network

• US 6,353,816 B1
• Filed: 06/22/1998
• Issued: 03/05/2002
• Est. Priority Date: 06/23/1997
• Status: Expired due to Fees

• Current Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation)
• Original Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation)
• Inventors: Tsukimoto, Hiroshi
• Primary Examiner(s): Davis, George B.
• Application Number:
  US09/100,493
• Time in Patent Office: 1,352 Days
• Field of Search: 706/13, 706/17, 706/21
• US Class Current: 706/17
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

A neural network analysis apparatus has an input means 1 for inputting multilinear functions that represent the various units of a trained neural network, a storage means 2 therefor, term generators 3 that successively generate terms of a Boolean function used for approximation from the coefficients of each term of the multilinear function input for each unit, a judgment condition storage section 22 into which is stored the conditions for generating each term of the Boolean function to be used in the approximation, an apparatus 5 that links the generated Boolean function terms, and a Boolean function synthesizing apparatus 6 for representing the output unit by input variables. The term generators 3 each have four sub-apparatuses: a data-limiting apparatus 31, a minimum value calculation apparatus 32, a non-important, attribute processing apparatus 33, and a function judgment apparatus 34.

Claims

1. A neural network analysis method comprising:

inputting each intermediate and output element of a trained neural network to be analyzed as data represented by a multilinear function, the neural network having been trained by learning data in a given domain;

approximating each intermediate and output element with a Boolean function approximation;

synthesizing the Boolean function approximation of each intermediate and output element into a synthesized Boolean function; and

outputting data, including the synthesized Boolean function, which is indicative of a predictive accuracy of the trained neural network, as an analysis result;

wherein approximating each intermediate and output element with a Boolean function approximation further comprises,

generating terms of a Boolean function for each intermediate and output element, and

linking the terms using a logical sum to obtain a Boolean function approximation of each intermediate and output element, and

said generating terms of a Boolean function for each intermediate and output element repeats a process comprising selecting a term that is made from a variable representing each intermediate and output element, limiting the learning data in the given domain to data within a limited subdomain corresponding to the term that is made from a variable, and making a judgment as to whether the term that is made from a variable exists in the Boolean function approximation based on data within the limited subdomain.

2. A neural network analysis method according to claim 1, wherein approximating each intermediate and output element with a Boolean function approximation and making a judgment comprises:

making a judgment as to whether a minimum value of each element, within the limited subdomain, exceeds a prescribed threshold value; and

making a judgment that the term that is made from a variable exists in the Boolean function approximation when the minimum value exceeds the prescribed threshold value.

3. A neural network analysis method according to claim 2, wherein approximating each intermediate and output element with a Boolean function approximation comprises:

establishing a judgment quantity for judging a size of each attribute of the learning data; and

approximating each element not using attributes for which the judgment quantity is at most a prescribed value and using only attributes for which the judgment quantity exceeds the prescribed value.

4. A neural network analysis method according to claim 3, wherein approximating each element with a Boolean function approximation comprises:

finding a minimum value of each element using an attribute for which the judgment quantity is largest;

extracting a Boolean function approximation based on the minimum;

calculating an error between the Boolean function approximation and the output value of each intermediate and output element and comparing the error with a given value;

when the error is above the given value, finding a minimum value of each intermediate and output element using another attribute for which the judgment quantity is a next largest attribute; and

repeating the steps of finding a minimum value of each element using another attribute, extracting a Boolean function approximation, and calculating and comparing an error with the given value, until the error becomes at least one of larger than the previous error and below the given value, so as to reduce the error between the Boolean function approximation and a value of each intermediate and output element.

5. A neural network analysis method according to claim 2, wherein approximating each element with a Boolean function comprises:

finding a minimum value of each element with regard to an attribute for which the judgment quantity exceeds the prescribed value; and

taking an average value of each element with regard to another attribute for which the judgment quantity does not exceed the prescribed value.

6. A neural network analysis method according to claim 1, wherein the approximating each element with a Boolean function approximation comprises:

approximating each element with a continuous Boolean function approximation that is a qualitative representation when the domain of the learning data is continuous.

7. A neural network analysis system comprising:

an input section for inputting each intermediate and output element of a neural network to be analyzed as data represented by a multilinear function, the neural network having been trained by learning data in a given domain;

a function-extracting section for approximating each input intermediate element and output element with a Boolean function approximation;

a function-synthesizing section for synthesizing the Boolean function approximation of each intermediate and output element obtained by each function-extracting section into a synthesized Boolean function; and

an output section for outputting, data including the synthesized Boolean function which is indicative of a predictive accuracy of the trained neural network as an analysis result;

wherein the function-extracting section comprises, a term-generating section for generating each term of the Boolean function for each intermediate and output element, and

a term-linking section for linking terms generated by the term-generating section using a logical sum to obtain Boolean function approximating, and

the term-generating section has a data-limiting section for limiting the learning data in the given domain to data within a limited subdomain corresponding to a term to be generated, and a judgment section for making a judgment as to whether the term exists in the Boolean function approximation based on data within the subdomain limited by the data-limiting section.

8. A neural network analysis system according to claim 7, wherein the judgment section of the term-generating section is configured to make a judgment as to whether a minimum value of each intermediate and output element, within the limited subdomain, exceeds a prescribed threshold value, and make a judgment that the term exists in the Boolean function approximation when the minimum value exceeds the prescribed threshold value.

9. A neural network analysis system according to claim 8, wherein the function-extracting section is configured to establish a judgment quantity for judging size of each attribute of the learning data, and approximate each intermediate and output element not using attributes for which the judgment quantity is at most a prescribed value and using only attributes for which the judgment quantity exceeds the prescribed value.

10. A neural network analysis system according to claim 9,

wherein the function-extracting section comprises an error-calculating section for calculating an error between a Boolean function approximation and the output value of each element obtained by the term-generating section and the term-linking section; and

the function-extracting section is configured to repeat the extraction of a Boolean function approximation by the term-generating section and the term-linking section and the calculation of an error by the error-calculating section, using attributes in decreasing order of the size of the judgment quantity, until the error calculated by the error-calculating section becomes at least one of larger than the previous error and below the given value, so as to reduce the error between the Boolean function approximation each of the intermediate and output and a value of the element.

11. A neural network analysis system according to claim 8, wherein the function-extracting section is configured to find a minimum value of each element with regard to an attribute for which the judgment quantity exceeds the prescribed value, and take an average value of each element with regard to another attribute for which the judgment quantity does not exceed the prescribed value.

12. A neural network analysis method according to claim 7, wherein the function-extracting section is configured to approximate each intermediate and output element with a continuous Boolean function approximation that is a qualitative representation when the domain of the learning data is continuous.

13. A storage medium onto which is stored a computer program, wherein the program comprises:

a means for causing a computer to input each intermediate and output element of a neural network to be analyzed as data represented by a multilinear function, the neural network having been trained by learning data in a given domain;

a means for causing a computer to generate each term of a Boolean function by repeating a process comprising selecting a term that is made from a variable representing each intermediate element and output element, limiting the learning data in the given domain to data within a subdomain corresponding to the term, making a judgment as to whether the term that is made from a variable exists in Boolean function approximation based on data within the limited subdomain;

a means for causing a computer to approximate each intermediate and output element with a Boolean function approximation by linking generated terms using a logical sum to obtain the Boolean function approximation;

a means for causing a computer to synthesize the Boolean function approximation of each intermediate and output element into a synthesized Boolean function; and

a means for causing a computer to output data, including the synthesized Boolean function, which is indicative of a predictive accuracy of the trained neural network as an analysis result.