Method and apparatus for adaptive classification

• US 5,479,574 A
• Filed: 04/01/1993
• Issued: 12/26/1995
• Est. Priority Date: 04/01/1993
• Status: Expired due to Term

• Current Assignee: American Traffic Solutions Incorporated (Verra Mobility Corp.)
• Original Assignee: Nestor, Inc.
• Inventors: Scofield, Christopher L., Nunez, Linda I. Mensinger, Glier, Michael T.
• Primary Examiner(s): Downs, Robert W.
• Application Number:
  US08/042,232
• Time in Patent Office: 999 Days
• Field of Search: 395/23, 395/27
• US Class Current: 706/25
• CPC Class Codes: G06F 18/24143 Distances to neighbourhood prototypes, e.g. restricted Coulomb energy networks [RCEN]
• CPC Class Codes: G06N 3/063 using electronic means
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

A neural network and method for pipeline operation within a neural network which permits rapid classification of input vectors provided thereto is disclosed. In a training mode, a plurality of training input features are presented to the neural network and distances between the plurality of training features and a plurality of prototype weight values are concurrently computed. In response to an indication of a last training epoch count values for each of the prototype weight values are stored in a memory to thereby allow the neural network to operate in a probabilistic classification mode.

Claims

1. A method of training a neural network having an input layer, a middle layer and an output layer, comprising the steps of:

(a) presenting an input vector having a plurality of training features to the neural network;

(b) computing distances between a plurality of said training features and a plurality of prototype weight values;

(c) generating, for each prototype weight value, a count value corresponding to a number of occurrences of an input vector that falls within a region of influence of a particular prototype;

(d) repeating steps (a)-(c) until the neural network provides an indication of a last training epoch; and

(e) in response to the indication of the last training epoch, storing, in a memory, the count value for each of the prototype weight values.

2. The method of claim 1 wherein the step of storing the count value corresponding to the number of occurrences of an input vector within a prototype includes the step of storing the count value in a memory location corresponding to a register for the storage of a middle layer weight value.

3. The method of claim 2 wherein the count value corresponds to a scalar value.

4. The method of claim 1 wherein the distances computed in said computing step are computed concurrently.