Neural network with non-linear transformations

• US 4,979,126 A
• Filed: 03/30/1988
• Issued: 12/18/1990
• Est. Priority Date: 03/30/1988
• Status: Expired due to Term

• Current Assignee: Computer Associates Think Inc. (Broadcom, Inc.)
• Original Assignee: AI Ware, Inc. (Broadcom, Inc.)
• Inventors: Khatibi, Farrokh, Pao, Yoh-Han
• Primary Examiner(s): MacDonald, Allen
• Application Number:
  US07/175,184
• Time in Patent Office: 993 Days
• Field of Search: 382/14, 382/15, 364/200, 364/900, 364/300, 364/513
• US Class Current: 706/25
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology

Abstract

A neural network system includes means for accomplishing artificial intelligence functions in three formerly divergent implementations. These functions include: supervised learning, unsupervised learning, and associative memory storage and retrieval. The subject neural network is created by addition of a non-linear layer to a more standard neural network architecture. The non-linear layer functions to expand a functional input space to a signal set including orthonormal elements, when the input signal is visualized as a vector representation. An input signal is selectively passed to a non-linear transform circuit, which outputs a transform signal therefrom. Both the input signal and the transform signal are placed in communication with a first layer of a plurality of processing nodes. An improved hardware implementation of the subject system includes a highly parallel, hybrid analog/digital circuitry. Included therein is a digitally addressed, random access memory means for storage and retrieval of an analog signal.

Claims

1. A neural network system comprising:

a plurality of processing nodes arranged into at least first and second layers;

input means adapted for receiving an input signal comprising data representative of an input pattern;

first communication means for communicating at least a portion of the input signal to at least a first portion of the processing nodes of the first layer;

transformation means for performing a non-linear transform on a signal input thereto to form a transformed signal as an output thereof;

second communication means for communicating at least a portion of the input pattern to the transformation means;

third communication means for communicating the transformed signal resultant from operation of the transform means to at least a second portion of the processing nodes of the first layer uniquely defined from the processing nodes of the first portion;

fourth communication means for communicating the output of processing nodes of the first portion and of the second portion of the first layer to at least a first portion of the processing nodes of the second layer; and

means for outputting the output signal from the processing nodes of the second layer.

2. The neural network of claim 1 wherein the second communication means includes means for communicating all portions of the input signal to the transformation means.

3. The neural network of claim 2 wherein the input pattern data is representative of vector data, and wherein the transformation means includes means for performing vector operations on the input data.

4. The neural network of claim 3 wherein the means for performing vector operations includes means for calculating an outer product of data representative of an input pattern.

5. The neural network of claim 1 further comprising pruning means for selectively pruning outputs of the transform means.

6. The neural network of claim 5 further comprising:

means for receiving data representative of a training set pattern;

evaluating means for evaluating joint activations of nodes over a training set pattern; and

means for controlling the pruning means in accordance with an output of the evaluating means.

7. The neural network of claim 1 wherein each processing node includes:

means for receiving an analog signal;

means for storing a digital value representative of a signal weighing value; and

means for performing an analog processing of a signal communicated thereto in accordance with the signal weighing value.

8. A method of neural processing comprising the steps of:

receiving an input signal comprising data, representative of an input pattern, into a plurality of processing nodes arranged into at least first and second layers;

communicating at least a portion of the input signal to at least a first portion of the processing nodes of the first layer;

communicating at least a second portion of the input signal to a transformation means;

performing, in the transform means, a non-linear transform on at least the second portion of the input signal to form a transform signal;

communicating the transform signal to at least a second portion of the processing nodes of the first layer uniquely defined from the processing nodes of the first portion;

communicating the output of processing nodes of the first portion and the second portion of the first layer to at least a first portion of the processing nodes of the second layer; and

outputting an output signal from the processing nodes of the second layer.

9. The method of claim 8 further comprising the step of communicating all portions of the input signal to the transformation means.

10. The method of claim 9 wherein the input pattern data is representative of vector data, and further comprising the step of performing vector operations on the input data.

11. The method of claim 10 further comprising the step of calculating an outer product of data representative of an input pattern during the step of performing vector operations on the input data.

12. The method of claim 11 further comprising the step of pruning selected outputs of the transform means.

13. The method of claim 12 further comprising the steps of:

receiving data representative of a training set pattern;

evaluating joint activations of nodes over a training set pattern and;

performing the pruning step in accordance with the evaluating.

14. A neural network system comprising:

a plurality of processing nodes arranged into first and second layers;

input means adapted for receiving an input signal comprising data representative of an input pattern;

first communication means for communicating at least a portion of the input signal to a first subset of the processing nodes of the first layer;

transformation means for performing a non-linear transform on a signal input thereto to form a transformed signal as an output thereof;

second communication means for communicating at least a portion of the input pattern to the transformation means;

third communication means for communicating the transformed signal resultant from operation of the transform means to at least a second subset of the processing nodes of the first layer, the second subset being uniquely defined from processing nodes of the first subset;

fourth communication means for selectively communicating the output of processing nodes of the first subset and of the second subset of the first layer to selected processing nodes of the second layer; and

means for outputting the output signal from the processing nodes of the second layer.

15. The neural network of claim 14 wherein each processing node includes:

means for receiving an analog signal;

means for storing a digital value representative of a signal weighing value; and

means for performing an analog processing of a signal communicated thereto in accordance with the signal weighing value.