Neural networks with Hebbian learning do not outperform random ones in fan-out systems

摘要

There are many network structures in the brain that have a strong divergence of connections from one processing layer to the next. Good examples are found in the olfactory system of insects and in the Hippocampus. This paper elaborates on the advantages or disadvantages of using Hebbian learning in fan-out systems as opposed to sparse randomly connected networks. The measure to determine performance is information conservation from one layer to the next. We conclude that in fan-out systems Hebbian learning does not improve information conservation and increases the level of activity in the network.