The effect of synaptic disconnection on bi-directional associative recall

The effects of synaptic disconnection ("damage") on associative recall in bi-directional associative memories (BAMs) are reported. The degradation of recall under various types of damage are described. The extent of failure is seen to be significantly dependent on the nature as well as the extent of damage caused to the BAM. A BAM is a two-level nonlinear neural network that recalls associative pairs of bit patterns. Particularly in the presence of noise. Like other neural networks, BAMs are fault tolerant, since they are distributed parallel processing elements, with each node contributing to the final output response. The paper describes the basic coding and decoding strategies as well as the strategies followed for "damaging" the BAM. In the experiments reported, graceful degradation, rather than catastrophic failure was observed. The results show that the recall is most sensitive to clustered synaptic disconnection and least sensitive to tandem synaptic disconnection. Some similarities to biological systems are mentioned.<
>