Encouraging networks modularity by seeding motifs

We propose a motifs seeding method to encourage the emergence of modular structure during network evolution. Previous studies fail to trigger modularity on freeform evolving ANNs either when varying environmental factors or the evolutionary process itself. We extracted statistical profiles of 3-node and 4-node motifs from evolved networks, and then generated new networks by seeding the most useful 3-node motif (feed-forward loop, ID:38). A series of retina recognition experiments was conducted using the seeded networks. The performance of different algorithms was measured. Our results indicate that modularity could be encouraged under certain conditions. We were able to build networks meeting a desired Z-score.