High Capacity Associative Memory Models with Bipolar and Binary, Biased Patterns

The high capacity associative memory model is interesting due to its significantly higher capacity when compared with the standard Hopfield model. These networks can use either bipolar or binary patterns, which may also be biased. This paper investigates the performance of a high capacity associative memory model trained with biased patterns, using either bipolar or binary representations. Our results indicate that the binary network performs less well under low bias, but better in other situations, compared with the bipolar network.