Redwood Neuroscience
Title "Attractor Neural Networks with
Depressing Synapses”
Bert
Kappen
The
Abstract:
We examine the role of depressing synapses on memory
retrieval in attractor neural networks. Using a standard mean field approach,
we analytically compute the phase diagram for small number of patterns as a
function of the noise and amount of depression of the synapses. Our results
demonstrate the appearance of a novel phase characterized by rapid switching
from one memory state to another. Subsequently, we show that the capacity of the
network to store patterns as fixed points of the attractor dynamics is strongly
reduced with depressing synapses. We confirm the validity of our mean field
results with numerical simulations. In addition, we argue to change the
traditional view of memories as stable fixed point attractors and suggest the
idea of dynamical memory as a possible mechanism of the brain.