Adaptive classifier having multiple subnetworks

• US 5,701,398 A
• Filed: 07/19/1996
• Issued: 12/23/1997
• Est. Priority Date: 07/01/1994
• Status: Expired due to Term

• Current Assignee: ACI Worldwide Corporation (ACI Worldwide Incorporated)
• Original Assignee: Nestor, Inc.
• Inventors: Cole, John, Laird, Mark, Glier, Michael T.
• Primary Examiner(s): Hafiz, Tariq R.
• Application Number:
  US08/684,683
• Time in Patent Office: 522 Days
• Field of Search: 382/155-159, 395/20-27
• US Class Current: 706/41
• CPC Class Codes: G06F 18/2414 Smoothing the distance, e.g. radial basis function networks [RBFN]
• CPC Class Codes: G06F 18/24143 Distances to neighbourhood prototypes, e.g. restricted Coulomb energy networks [RCEN]
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/063 using electronic means

Abstract

A neural network including a plurality of sub-nets stored in a memory array and a method of operation. Each of the sub-nets includes a corresponding plurality of weights and is individually operable to classify an input vector. A computation unit, including a distance calculation unit and a math unit, is responsive to an input vector comprising input training features for determining a distance between the weights of each sub-net to each of the input features of the input vector and for determining whether the distance is within a particular region of influence. Also described is a parallel process for training the sub-nets.

Claims

1. A neural network for processing an input pattern in an effort to classify said input pattern into one or more classes, said neural network comprising:

a memory array;

a controller operative to store a plurality of sub-nets in said memory array, each of said sub-nets comprising a corresponding plurality of weights;

a distance calculation unit coupled to said memory array for calculating a distance between at least one feature of said input pattern and said plurality of weights of one of said plurality of sub-nets;

a math unit coupled to said distance calculation unit for comparing said calculated distance with a region of influence value associated with each of said plurality of weights of said one of said plurality of sub-nets to provide a class response output signal, wherein said input pattern is classified into said one or more classes using said one of said plurality of sub-nets independently of other ones of said plurality of sub-nets; and

a prototype parameter memory for storing said region of influence value associated with each of said plurality of weights of said one of said plurality of sub-nets.

2. The neural network recited in claim 1 wherein said class response output signal is indicative of which, if any, of said one or more classes is firing.

3. The neural network recited in claim 1 wherein said class response output signal is indicative of a probability that said input pattern is associated with said one or more classes as computed by a probability density function.

4. The neural network recited in claim 1 wherein said class response output signal provides an input pattern for classification by another one of said plurality of sub-nets.

5. The neural network recited in claim 1 wherein said prototype parameter memory additionally stores a class type indicator.

6. The neural network recited in claim 1 wherein said prototype parameter memory additionally stores a smoothing factor.

7. The neural network recited in claim 1 wherein said distance calculation unit comprises a plurality of computation units, each one associated with at least one of said plurality of sub-nets, for simultaneously computing a distance between at least one feature of said input pattern and said plurality of weights of said at least one associated sub-net.

8. The neural network recited in claim 1 wherein said neural network is formed on an integrated circuit.

9. A neural network for processing an input pattern in an effort to classify said input pattern into one or more classes, said neural network comprising:

a prototype array distance calculation unit comprising;

a memory array comprising a plurality of memory array regions; and

a distance calculation unit comprising a plurality of computation units, each one associated with and coupled to at least one of said plurality of memory array regions;

a controller operative to store a plurality of sub-nets in said memory array, each sub-net comprising a corresponding plurality of weights and being stored in a respective one of said plurality of memory array regions;

a math unit responsive to at least one of said plurality of computation units for providing a class response output signal; and

a prototype parameter memory coupled to said math unit for storing a region of influence value associated with each of said plurality of weights of one of said plurality of sub-nets, wherein said math unit compares said calculated distance provided by one of said plurality of computation units associated with said one of said plurality of sub-nets with said region of influence value associated with each of said plurality of weights of said one of said plurality of sub-nets in order to provide said class response output signal.

10. The neural network recited in claim 9 wherein each of said plurality of computation units computes a distance between at least one input feature of said input pattern and a plurality of weights of a sub-net stored in said associated memory array region and wherein said math unit is responsive to said distance computed by said at least one computation unit for providing said class response output signal.

11. The neural network recited in claim 10 wherein each of said plurality of computation units computes said distance substantially simultaneously.

12. The neural network recited in claim 9 wherein said prototype parameter memory additionally stores a class type indicator.

13. The neural network recited in claim 9 wherein said prototype parameter memory additionally stores a smoothing factor.