Cortronic neural networks with distributed processing

• US 6,366,897 B1
• Filed: 07/26/1999
• Issued: 04/02/2002
• Est. Priority Date: 07/26/1999
• Status: Expired due to Term

• Current Assignee: Fair Isaac Corporation
• Original Assignee: HNC Software Inc. (Fair Isaac Corporation)
• Inventors: Calmbach, Richard, Means, Robert W.
• Primary Examiner(s): Davis, George B.
• Application Number:
  US09/361,776
• Time in Patent Office: 981 Days
• Field of Search: 706/27, 706/39, 706/41
• US Class Current: 706/27
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

A cortronic neural network defines connections between neurons in a number of regions using target lists, which identify the output connections of each neuron and the connection strength. Neurons are preferably sparsely interconnected between regions. Training of connection weights employs a three stage process, which involves computation of the contribution to the input intensity of each neuron by every currently active neuron, a competition process that determines the next set of active neurons based on their current input intensity, and a weight adjustment process that updates and normalizes the connection weights based on which neurons won the competition process, and their connectivity with other winning neurons.

Claims

1. A distributed artificial neural network system, comprising:

a plurality of distributed computers, each distributed computer storing in memory,

at least one region of an artificial neural network, each region containing a plurality of the neurons, each neuron associated with a list of target neurons to which the neuron has an output connection, the output connection having a connection weight, and wherein for at least one region on a distributed computer, the neurons in the region being connected to substantially less than all of the neurons in another region or the same region;

an intensity calculation process, executed by the distributed computer, that determines an intensity of inputs to each neuron from the connection weights of the output connections of other neurons for which the neuron is a target neuron;

a competition calculation process, executed by the distributed computer, that determines a subset of the neurons that are active neurons as a function of an intensity of inputs at each neuron; and

a renormalization calculation process, executed by the distributed computer, that normalizes the connection weights of the output connections of a neuron to its target neurons by a scaling factor derived from the sum of the connection weights of the output connections, and that normalizes the input intensity of each neuron by a scaling factor derived from the sum of input connection weights of connections from other neurons.

2. The artificial neural network system of claim 1, wherein the intensity calculation process on each distributed computer determines the intensity of inputs to

each neuron from the weights that each neuron has to each of its target neurons by:

accumulating for each neuron, the weights of connections that local neurons on the distributed computer have to the neuron;

providing the accumulated weights to each of the other distributed computers;

receiving the accumulated weights computed by other distributed computers; and

combining the received accumulated weights.

3. A computer system for operating a distributed artificial neural network, comprising:

an executive computer storing a region map defining a plurality of regions, each region including a plurality of neurons, and mapping a global identifier of each neuron to a local identifier of the neuron in a region; and

a plurality of distributed computers communicatively coupled to the executive computer and to each other, each distributed computer storing the neurons of one region or portion thereof, each neuron associated with a list of target neurons, with each target neuron receiving an output from the neuron, the output having a connection weight, wherein the target neuron is in the same region as the neuron, or in a different region;

each distributed computer further comprising;

an intensity calculation process that determines an intensity of inputs to each neuron from the connection weights of the output connections of other neurons for which the neuron is a target neuron;

a competition calculation process that determines a subset of the neurons that are active neurons as a function of an intensity of inputs at each neuron; and

a renormalization calculation process that normalizes the connection weights of the output connections of a neuron to its target neurons by a scaling factor derived from the sum of the connection weights of the output connections, and that normalizes the input intensity of each neuron by a scaling factor derived from the sum of input connection weights of connections from other neurons.