Neural network shell for application programs

• US 5,142,665 A
• Filed: 02/20/1990
• Issued: 08/25/1992
• Est. Priority Date: 02/20/1990
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Bigus, Joseph P.
• Primary Examiner(s): MacDonald, Allen R.
• Application Number:
  US07/482,450
• Time in Patent Office: 917 Days
• Field of Search: 364/513, 395/21, 395/23, 395/76, 395/26, 395/22, 395/11
• US Class Current: 706/44
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology

Abstract

A neural network shell has a defined interface to an application program. By interfacing with the neural network shell, any application program becomes a neural network application program. The neural network shell contains a set of utility programs that transfers data into and out of a neural network data structure. This set of utility programs allows an application program to define a new neural network model, create a neural network data structure, train a neural network, and run a neural network. Once trained, the neural network data structure can be transported to other computer systems or to application programs written in different computing languages running on similar or different computer systems.

Claims

1. A method in a neural network shell of enabling an application program to run a neural network model, thereby becoming a neural network having an input for data and an output for a result, said method comprising the machine-executed steps of:

creating a neural network data structure, said creating step further comprising the steps of;

selecting a neural network model from a plurality of defined neural network models for said neural network, said neural network model having a corresponding default neural network data structure;

initializing said corresponding default neural network data structure for said neural network model selected in said selecting step by inputting initial values for a plurality of data arrays into said default neural network data structure;

prompting for a plurality of parameters specific to said neural network;

inputting said plurality of parameters into said default neural network data structure;

teaching said neural network, said teaching step further comprising the steps of;

presenting training data at said input of said neural network;

repeatedly adjusting the values of said plurality of data arrays until said result at said output is within tolerance of a correct result;

locking the values of said plurality of data arrays responsive to said adjusting step;

running said neural network, said running step further comprises the steps of;

presenting actual data at said input of said neural network; and

retrieving the result from the output of said neural network.