Hybrid neural network classifier, systems and methods

• US 5,943,661 A
• Filed: 07/11/1991
• Issued: 08/24/1999
• Est. Priority Date: 07/11/1991
• Status: Expired due to Term

• Current Assignee: Intel Corporation
• Original Assignee: Texas Instruments, Inc.
• Inventors: Katz, Alan Jerry
• Primary Examiner(s): Hafiz, Tariq R.
• Application Number:
  US07/728,426
• Time in Patent Office: 2,966 Days
• Field of Search: 395/51, 395/22, 395/26, 395/24, 395/25, 395/23, 395/13, 395/934, 364/413.05, 706/15-18, 706/23-25
• US Class Current: 706/16
• CPC Class Codes: G06F 18/2321 using statistics or function optimisation, e.g. modelling of probability density functions
• CPC Class Codes: G06F 18/2414 Smoothing the distance, e.g. radial basis function networks [RBFN]
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 3/105 Shells for specifying net layout

Abstract

A method of and system for parallelizing an program, comprising the steps of inputting an algorithm, operating said algorithm on selected data inputs to generate representative outputs, inputting representative outputs into parallelizing algorithms, and outputting a parallel implementation of said algorithm. In particular, this provides a parallel framework for target classification and pattern recognition procedures.

Claims

1. A method of training a neural network classifier, comprising the steps of:

(a) providing a first set of target points Z₁, Z₂, . . . Z_L in a feature space;

(b) forming an estimated target probability density P on said feature space from said target points Z₁, Z₂, . . . Z_L ;

(c) providing a second set of target points W₁, W₂, . . . W_M in said feature space;

(d) defining a threshold T from the number of W_j with P(W_j)>

T and the number of W_j with P(W_j)<

T;

(e) providing a third set of points X₁, X₂, . . . X_N in said feature space, and forming a set of pairs (X_j, Y_j) where Y_j is "target" when P(X_j)>

T and Y_j is "clutter" when P(X_j)<

T; and

(f) using the pairs (X₁, Y₁), (X₂, Y₂), . . . , (X_j, Y_j), . . . , (X_N, Y_N) as input/output pairs to train a neural network classifier.

2. The method of claim 1, wherein:

(a) said step (b) of forming an estimated target probability density includes compressing said set of target points Z₁, Z₂, . . . Z_L.

3. The method of claim 1, wherein:

(a) said step (f) of using pairs to train a neural network classifier includes backpropagation in said neural network.

4. The method of claim 1, wherein:

(a) said step (b) of forming an estimated target probability density includes picking a probability density kernel k(Z) defined on said feature space and summing the translates of said kernel, k(Z-Z_j), for all Z_j in said first set of target points.