Extreme pattern compression in lognormal networks
Gabriele Scheler · 2016
Traditional neural networks are constructed with random or sparse connectivity and with uniform neurons, both excitatory and inhibitory. Biological neural networks have a power-law (lognormal) distribution of weights, spike frequency, and intrinsic excitability. We constructed networks with these properties and stimulated them with rate-coded patterns to selected input neurons. The goal was to produce expanded representations for input on a lognormal neural field and to analyze the properties of these representations.It has been noticed in cortical regions as well as in hippocampal CA1/CA3 that responses to behavioral stimulation affect 20-40% of neurons, and that a small number of neurons (1-3%) are sufficient to uniquely recognize which pattern/event is being processed. We were able to reproduce both results with a biological (lognormal) neural network. We found that lognormal neural networks are highly efficient and precise. LG networks need less spikes and less synaptic weights, yet the high variability in spike response makes it easier to recognize patterns.