Attractor networks in systems with underlying random connectivity
P.E. Latham
Most treatments of Hopfield networks [1] use a deterministic weight
matrix. We consider the more realistic case in which the weight
matrix has an underlying random component associated with sparse
connectivity and heterogeneous coupling. Randomly connected networks
of excitatory and inhibitory neurons (i.e., networks with no
structure) exhibit, over a broad range of parameters, a single stable
state at low firing rate. We investigate the conditions for the
formation of additional fixed points -- new memories -- as
connectivity among subpopulations of neurons increases. Using both
mean field analysis and simulations with spiking model neurons, we
find that there is a threshold connectivity for memories to form, that
memories form at high firing rate, and that memories can form
only if
the neuronal transfer function -- the firing rate of a post-synaptic
neuron as a function of the mean firing rates of the pre-synaptic
neurons coupled to it -- has a particular shape.
1. J.J. Hopfield, PNAS, 1982.