Apparatus and method for facilitating use of a neural network

• US 5,317,676 A
• Filed: 12/03/1992
• Issued: 05/31/1994
• Est. Priority Date: 04/18/1991
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Bigus, Joseph P., Austvold, Shawn M., Henckel, Jonathan D., Hospers, Paul A.
• Primary Examiner(s): MacDonald, Allen R.
• Application Number:
  US07/986,889
• Time in Patent Office: 544 Days
• Field of Search: 395/22, 395/23, 395/24, 395/27, 395/76, 395/26, 364/715.03, 364/715.05
• US Class Current: 706/44
• CPC Class Codes: G06N 3/105 Shells for specifying net layout

Abstract

A neural network development utility assists a developer in generating one or more filters for data to be input to or output from a neural network. A filter is a device which translates data in accordance with a data transformation definition contained in a translate template. Source data for the neural network may be expressed in any arbitrary combination of symbolic or numeric fields in a data base. The developer selects those fields to be used from an interactive menu. The utility scans the selected field entries in the source data base to identify the logical type of each field, and creates a default translate template based on this scan. Numeric data is automatically scaled. The developer may use the default template, or edit it from an interactive editor. When editing the template, the developer may select from a menu of commonly used neural network data formats, and from a menu of commonly used primitive mathematical operations. The developer may interactively define additional filters to perform data transformations in series, thus achieving more complex mathematical operations on the data. Templates may be edited at any time during the development process. If a network does not appear to be giving satisfactory results, the developer may easily alter the template to present inputs in some other format.

Claims

1. An apparatus for facilitating the use of a neural network, comprising:

means for receiving source data for the neural network, said source data comprising a plurality of source data fields;

translation definition means for defining, with respect to each of said source data fields, a corresponding data transformation, wherein at least one of said data transformations is a transformation of data of one data type to data of another data type, and wherein said translation definition means comprises means for automatically generating a default data transformation definition with respect to each of said source data fields; and

means, responsive to said translation definition means, for translating source data contained in each said source data field according to the data transformation corresponding to said source data field, to produce a neural network input.

2. The apparatus for facilitating the use of a neural network of claim 1, further comprising:

means for receiving output data from the neural network, said output data comprising at least one output data field;

output translation definition means for defining, with respect to an output data field, a corresponding data transformation of data of one data type to data of another data type; and

means, responsive to said output translation definition means, for translating output data contained in said output data field according to the data transformation corresponding to said output data field, to produce a translated neural network output.

3. The apparatus for facilitating the use of a neural network of claim 1, wherein said means for automatically generating a default data transformation definition comprises means for automatically defining a scaling function with respect to each numeric source data field for translating source numeric data to numeric data in a predetermined range of values.

4. An apparatus for facilitating the use of a neural network, comprising:

means for receiving source data for the neural network, said source data comprising a plurality of source data fields;

translation definition means for defining, with respect to each of said source data fields, a corresponding data transformation, wherein at least one of said data transformations is a transformation of data of one data type to data of another data type, wherein said translation definition means comprises;

(a) means for generating a translate template defining, with respect to each of said source data fields, a corresponding data transformation;

(b) means for storing and retrieving said translate template; and

(c) means for editing said translate template after it has been initially generated; and

means, responsive to said translation definition means, for translating source data contained in each said source data field according to the data transformation corresponding to said source data field, to produce a neural network input.

5. The apparatus for facilitating the use of a neural network of claim 4, wherein said translation definition means comprises means for automatically generating a default translate template.

6. The apparatus for facilitating the use of a neural network of claim 5, wherein said means for automatically generating a default translate template comprises:

means for scanning said source data to determine a logical data type for each of said source data fields;

means for associating a default target logical data type with each logical data type for a source data field; and

means for recording in said default translate template, with respect to each of said source data fields, said logical data type for the source data field and said default target logical data type associated with said logical data type for the source data field.

7. The apparatus for facilitating the use of a neural network of claim 5, wherein said means for automatically generating a default translate template comprises means for automatically defining a scaling function with respect to each numeric source data field for translating source numeric data to numeric data in a predetermined range of values.

8. A method for using a neural network, comprising the steps of:

receiving source data for the neural network, said source data comprising a plurality of source data fields;

defining, with respect to each of said source data fields, a corresponding data transformation, wherein at least one of said data transformations is a transformation of data of one data type to data of another data type, and wherein said step of defining a data transformation comprises;

(a) automatically scanning said source data to determine the content thereof, and (b) automatically generating a default data transformation definition with respect to each of said source data fields based on information obtained from said step of automatically scanning said source data; and

translating source data contained in each said source data field according to the data transformation corresponding to said source data field, to produce a neural network input.

9. The method for using a neural network of claim 8, further comprising the steps of:

receiving output data from the neural network, said output data comprising at least one output data field;

defining, with respect to an output data field, a corresponding data transformation of data of one data type to data of another data type; and

translating output data contained in said output data field according to the data transformation corresponding to said output data field, to produce a translated neural network output.

10. The method for using a neural network of claim 8, wherein said step of automatically generating a default data transformation definition comprises automatically defining a scaling function with respect to each numeric source data field for translating source numeric data to numeric data in a predetermined range of values.

11. A method for using a neural network, comprising the steps of:

receiving source data for the neural network, said source data comprising a plurality of source data fields;

defining, with respect to each of said source data fields, a corresponding data transformation, wherein at least one of said data transformations is a transformation of data of one data type to data of another data type, and wherein said step of defining a data transformation comprises generating a translate template defining, with respect to each of said source data fields, a corresponding data transformation; and

translating source data contained in each said source data field according to the data transformation corresponding to said source data field, to produce a neural network input.

12. The method for using a neural network of claim 11, further comprising the steps of:

storing said translate template;

retrieving said translate template from storage; and

editing said translate template after it has been initially generated.

13. The method for using a neural network of claim 11, wherein wherein said step of defining a data transformation comprises automatically generating a default translate template.

14. The method for using a neural network of claim 13, wherein said step of automatically generating a default translate template comprises:

scanning said source data to determine a logical data type for each of said source data fields;

associating a default target logical data type with each logical data type for a source data field; and

recording in said default translate template, with respect to each of said source data fields, said logical data type for the source data field and said default target logical data type associated with said logical data type for the source data field.

15. The method for using a neural network of claim 13, wherein said step of automatically generating a default translate template comprises:

scanning said source data to determine a range of source data values;

defining a scaling function with respect to each numeric source data field for translating source numeric data to numeric data in a predetermined range of values, said scaling function being dependent on said range of source data values; and

recording parameters of said scaling function in said default translate template.

16. A method for teaching a neural network having an input and an output for a result, comprising the steps of:

identifying source training data for said neural network;

defining a data transformation of said source training data, wherein said data transformation comprises transformation of data of one data type to data of another data type;

translating said source data in accordance with said data transformation definition to produce translated input training data;

presenting said translated input training data at said input of said neural network; and

repeatedly editing said data transformation definition until said result at said output is within tolerance of a correct result.

17. The method for teaching a neural network of claim 16, wherein said step of defining a data transformation comprises:

automatically scanning said source data to determine the content thereof; and

automatically generating a default data transformation definition with respect to each of said source data fields based on information obtained from said step of automatically scanning said source data.

18. The method for teaching a neural network of claim 17, wherein said step of automatically generating a default data transformation definition comprises automatically defining a scaling function with respect to each numeric source data field for translating source numeric data to numeric data in a predetermined range of values.

19. The method for teaching a neural network of claim 16, wherein:

said step of defining a data transformation comprises generating a translate template defining, with respect to each of said source data fields, a corresponding data transformation; and

said step of repeatedly editing said data transformation definition comprises repeatedly editing said translate template.

20. A neural network driven artificial intelligence apparatus for producing results from source data, comprising:

a neural network, having a numeric input of a pre-defined data type and a numeric output;

means for automatically translating said source data to numeric input for said neural network, wherein said source data comprises data of a type other than said pre-defined data type of said numeric input, and wherein said means for automatically translating said source data comprises;

(a) means for generating a translate template defining, with respect to each of a plurality of source data fields, a corresponding data transformation;

(b) means for storing and retrieving said translate template;

(c) means for editing said translate template after it has been initially generated; and

(d) means for translating source data according to the data transformation definition contained in said translate template; and

means for presenting said source data translated to numeric input to said neural network.

21. The neural network driven artificial intelligence apparatus of claim 20, wherein said neural network comprises:

a programmable processor;

storage means for storing programs and data accessible to said processor; and

a neural network simulation program, contained in said storage means, for executing on said programmable processor and thereby becoming a neural network.

22. The neural network driven artificial intelligence apparatus of claim 20, wherein said means for generating a translate template comprises:

means for scanning said source data to determine a logical data type for each of said source data fields;

means for associating a default target logical data type with each logical data type for a source data field; and

means for recording in said translate template, with respect to each of said source data fields, said logical data type for the source data field and said default target logical data type associated with said logical data type for the source data field.

23. A neural network data filter, comprising:

means for receiving source data for the neural network;

means for receiving a translate template, said translate template comprising a logical input field specifying a logical data type for a source data field, and a logical output field specifying a logical data type for an output; and

means for converting said source data from a form specified by said logical input field to a form specified by said logical output field.

24. The neural network data filter of claim 23, wherein:

said translate template further comprises a mathematical operator field specifying a mathematical operation; and

said filter further comprises means for performing the mathematical operation specified by said mathematical operator field on data being converted by said converting means.

25. The neural network data filter of claim 23, wherein:

said translate template further comprises a scaling parameters field specifying parameters of a scaling operation; and

said filter further comprises means for performing the scaling operation specified by said scaling parameters field on data being converted by said converting means.