Communication in a society of interacting perceptrons

A numerical investigation into the behavior of populations of interacting neural networks is presented. A number of neural networks with randomly generated weights interact with each other and share their knowledge about the same phenomenon through reducing their output differences. Results of this study show that all the interacting neural networks reside in the same state of knowledge, here defined as the state of common knowledge of the population, although their final vectors of connection weights would not necessarily be the same. Also segregation may take place at some stage during their interactions where clusters of neural networks with similar knowledge are formed. However these temporary clusters attract each other and dissolve into one dominant cluster representing the state of common knowledge of the population. Also it was observed that the final knowledge of the population is generally completely different from the average of the initial knowledge of the population. The neural networks used in this study have been three layered perceptrons of similar architecture and activation functions. Populations of 2, 20, 50 and 99 neural networks have been studied. No general conclusions about the subject can be drawn at this stage of the studies