Encoding Sparse Features in a Bidirectional Associative Memory

Bidirectional Associative Memories (BAMs) are artificial neural networks that can learn and recall various types of associations. Although BAM models have shown great promise at modeling human cognitive processes, these models have often been investigated under conditions where stimuli are densely represented using a bipolar coding scheme. However, research has shown that dense representations are energetically costly given that various stimulus representations need to be detected, processed and analyzed on a daily basis. Instead, biological networks work on minimizing energy expenditure by encoding sparse stimulus features that maximize information representation. This paper extends this line of search and shows that BAM models can improve learning and recall performance in a sparse encoding regime. It provides a strategy for artificial neural networks that seek to maintain valuable processing resources, especially under constraints of noisy representations of stimulus features.