Associative memory in a recurrent neural network


  1. Barton, S.A.
Corporate Authors
Defence Research Establishment Suffield, Ralston ALTA (CAN)
It is shown that unsupervised learning in a recurrent neural network (RNN) can lead to explicit associative memory. Using two separate sensor arrays during training, the RNN stores the information necessary to reproduce both images when presented with a single sensor image after training. This is a system that is capable of learning to make connections between different forms of sensory data. Memory is encoded in the strengths of the connection between the nodes of the RNN. Only patterns that have been seen during training generation strong recognition signals. Other patterns generate signals indicating that they can not be identified. The storage capacity of the RNN depends on the number of nodes and the number of their input connections. It is hsown that the RNN memory capacity does not decrease abruptly with increasing number of training patterns. In fact, the reproduction precision decays gradually. For a given RNN size, the memory simply becomes less clear as more storage is attempted. If unacceptable indications of recognition and identification are obtained, it is only necessary to increase the number of nodes. TRUNCATED
Unmanned vehicles;Machine intelligence;Autonomous machines;Unsupervised learning;Associative memory
Report Number
DRES-TM-2001-053 — Technical Memorandum
Date of publication
01 May 2001
Number of Pages
Hardcopy;Document Image stored on Optical Disk

Permanent link

Document 1 of 1

Date modified: