Embedding neural networks into spreadsheet applications

• US 5,241,620 A
• Filed: 01/03/1991
• Issued: 08/31/1993
• Est. Priority Date: 01/03/1991
• Status: Expired due to Fees

• Current Assignee: Promised Land Technologies Incorporated
• Original Assignee: Promised Land Technologies Incorporated
• Inventors: Ruggiero, Murray
• Primary Examiner(s): MacDonald, Allen R.
• Application Number:
  US07/637,301
• Time in Patent Office: 971 Days
• Field of Search: 395/22, 364/419
• US Class Current: 706/16
• CPC Class Codes: G06F 40/18 of spreadsheets form-filling G06F40/174
• CPC Class Codes: G06N 3/105 Shells for specifying net layout

Abstract

The present invention relates to a method of embedding a neural network into an application program such as a spreadsheet program. The method comprises providing an application program in which information is stored in rows and columns or a database containing fields and records and embedding a neural network in the application program or database using the stored information. The embedding step includes allocating unused memory in the application program and creating both a neural network engine and an application interface structure from the unused memory. Once the neural network engine and an application interface structure have been created, the neural network may be trained using variable numerical and symbolic data stored within the application program. Once training is completed, the neural network is ready for use, merely by using a recall function built into the applications program.

Claims

1. A method for incorporating a neural network into an application program and using said network, said method comprising:

providing an applications program in which information is stored;

embedding a neural network in said applications program using said stored information;

said embedding step comprising allocating unused memory in said applications program and creating a neural network engine and an application interface structure from said allocated unused memory; and

training said neural network utilizing variable numeric and symbolic data stored within said applications program, said training step including preparing said numeric variable data by storing in said interface structure for each numerical variable, a maximum possible value, a minimum possible value, a column location, a starting row location, information on the type of variable and a desired number of training patterns.

2. The method of claim 1 wherein said training step further includes:

preparing said variable symbolic data by storing in said interface structure for each piece of symbolic data, an address for the variable'"'"'s symbol table, a symbol type, a column location, a starting row location, information on the type of variable and said desired number of training patterns;

assigning a numerical value to each symbol in said table; and

storing the numerical value of the first symbol and the last symbol in the interface structure as the minimum and maximum.

3. The method of claim 2 wherein said training step further comprises:

(a) scaling all data to values between 0.0 and 1.0;

(b) performing one forward pass and one backward pass through the neural network; and

(c) repeating steps (a) and (b) for said desired number of training patterns.

4. The method of claim 3 wherein said training step further comprises:

(d) determining a percent error;

(e) comparing said percent error with a preset error to terminate training; and

repeating steps (a), (b), (c), (d), and (e) if said percent error is less than said preset error.

5. The method of claim 1 further comprising:

recalling said neural network after completion of said training;

inputting a desired number of inputs and a desired number of outputs; and

executing a forward pass through said trained neural network.

6. A method according to claim 1 wherein said training step includes updating weights in said neural network by reading values stored within an application memory.

7. A method according to claim 1 further comprising:

using error rate and outputted weight matrices to implement boosting of said training.

8. A method for incorporating a neural network into an application program and using said network, said method comprising:

providing an applications program in which information is stored;

embedding a neural network in said applications program using said stored information;

said embedding step comprising allocating unused memory in said applications program and creating a neural network engine and an application interface structure from said allocated unused memory;

training said neural network utilizing variable numeric and symbolic data stored within said applications program; and

solving any problem of excessive nodes by exporting a table of connection weights for said network using application command language;

examining said weight table for all connection weights;

converting connection weights close to zero to zero; and

importing said weight table with said converted connection weights back to said neural network.

9. A method for incorporating a neural network into an application program and using said network, said method comprising:

providing an applications program in which information is stored;

embedding a neural network in said applications program using said stored information;

said embedding step comprises allocating unused memory in said applications program and creating a neural network engine and an application interface structure from said allocated unused memory;

training said neural network utilizing variable numeric and symbolic data stored within said applications program;

recalling said neural network after completion of said training;

inputting a desired number of inputs and a desired number of outputs;

executing a forward pass through said trained neural network; and

said recalling step comprising

(a) specifying a number of columns;

(b) storing for each numeric column information about column location, starting row location, number of patterns in the data structure;

(c) storing for each symbolic column information about column location, starting row location, number of patterns, address of the symbol table in the data structure; and

(d) substituting in the application program, a symbolic value from the symbolic table with a number assigned to that symbol so that all columns are converted to numeric values.

10. The method of claim 9 further comprising:

scaling all numeric data to a number in the range of 0.0 to 1.0; and

using said scaled data for said forward pass through the neural network so as to produce network outputs in a range of from 0.0 to 1.0.

11. The method of claim 10 further comprising:

converting said outputs into unscaled numeric values; and

converting symbolic data from a numeric form to a symbolic form using an interface code.

12. A method according to claim 9 further comprising automatically repeating said recalling step whenever variable ones of said inputs are changed in an area defined by said interface structure by a user.

13. A method according to claim 9 further comprising automatically repeating said recalling step whenever a change in size of an area defined by the interface structure occurs.