Spreading activation and sparseness in a bidirectional associative memory

The Bidirectional Associative Memory (BAM) is a type of artificial neural network that was shown to bear great performances in learning and recalling various types of associations. However, this model has always been investigated under optimal conditions in which all the patterns have the same desirability and the network is fully connected. In this paper, the influence of spreading-activation and sparseness in a BAM network is studied. Results show that even under such variability the performances of the BAM are unaffected. This study gives us a better understanding of how attractors can be developed and could lead to more robust computational intelligence systems.