Topology influences performance in the associative memory neural networks
Jianquan Lu,Juan He,Jinde Cao,Zhiqiang Gao
DOI: https://doi.org/10.1016/j.physleta.2006.01.085
IF: 2.707
2006-01-01
Physics Letters A
Abstract:To explore how topology affects performance within Hopfield-type associative memory neural networks (AMNNs), we studied the computational performance of the neural networks with regular lattice, random, small-world, and scale-free structures. In this Letter, we found that the memory performance of neural networks obtained through asynchronous updating from “larger” nodes to “smaller” nodes are better than asynchronous updating in random order, especially for the scale-free topology. The computational performance of associative memory neural networks linked by the above-mentioned network topologies with the same amounts of nodes (neurons) and edges (synapses) were studied respectively. Along with topologies becoming more random and less locally disordered, we will see that the performance of associative memory neural network is quite improved. By comparing, we show that the regular lattice and random network form two extremes in terms of patterns stability and retrievability. For a network, its patterns stability and retrievability can be largely enhanced by adding a random component or some shortcuts to its structured component. According to the conclusions of this Letter, we can design the associative memory neural networks with high performance and minimal interconnect requirements.