TECHNIQUES FOR PATTERN CLASSIFICATION USING A CONVERGENT RECURRENT NEURAL NETWORK

PDF

Authors
  1. Barton, S.A.
Corporate Authors
Defence Research Establishment Suffield, Ralston ALTA (CAN)
Abstract
A recurrent neural network (RNN) using a structure and learning rules that have previously been shown to result in autonomous goal-driven motion in a simulated mobile machine, is now shown to be capable of memory, recognition and identification of patterns. Two techniques are presented. The first uses implicity memory, in which no direct representation of an image is stored, and an image may be identified by the point to which the RNN converges during exposure to the image. Data compression may be achieved by projecting the output of all the nodes in the RNN onto a vector that is of lower dimension than the image. This vector can be used to identify trained patterns. It is also possible to store and regenerate explicit representations of the actual images used in training. To achieve this, the learning rules used in the first technique wewre augmented by a resonant feedback signal that adjusts the input connection strengths of nodes in a memory region of the RNN. An image that has been used in training then stimulates reproduction of the same image in an output array coupled to the RNN memory region, and generates a signal indicating recogition. An image that has not been senn before, generates no reproduction or recognition signal. Based on the technique for explicit memory, designs for a self-learning associative memory are presented. TRUNCATED
Keywords
Countermine operations;Image classification;Unmanned vehicles;Remote Minefield Detection (RMD)
Report Number
DRES-709 —
Date of publication
01 Dec 1998
Number of Pages
55
DSTKIM No
99-00422
CANDIS No
510480
Format(s):
Hardcopy;Document Image stored on Optical Disk

Permanent link

Document 1 of 1

Date modified: