Structured connectivity in an associative memory model

High capacity associative memory models with dilute structured connectivity are trained using naturalistic bitmap patterns. The connectivity of the model is chosen to reflect the local spatial continuity of the data. The results show that the structured connectivity gives the networks a higher effective capacity than equivalent randomly diluted networks. Moreover the locally connected networks have a much lower mean connection length than the randomly connected network. It is shown that a small amount of additional connectivity can correct any neurons that fail to train.