Paradoxical increase of capacity due to spurious overlaps in attractor networks

TL;DR

This paper reveals that in certain associative memory models, spurious overlaps can paradoxically enhance storage capacity by reducing mean synaptic inputs, linking neurobiological plasticity to network performance.

Contribution

It demonstrates that neurobiologically inferred learning rules cause spurious overlaps to increase capacity by decreasing mean synaptic inputs, a novel insight into associative memory models.

Findings

Spurious overlaps reduce mean synaptic inputs in models with biologically inspired learning.

This reduction in mean inputs leads to increased storage capacity.

The study links plasticity rules to network capacity enhancement.

Abstract

In Hopfield-type associative memory models, memories are stored in the connectivity matrix and can be retrieved subsequently thanks to the collective dynamics of the network. In these models, the retrieval of a particular memory can be hampered by overlaps between the network state and other memories, termed spurious overlaps since these overlaps collectively introduce noise in the retrieval process. In classic models, spurious overlaps increase the variance of synaptic inputs but do not affect the mean. We show here that in models equipped with a learning rule inferred from neurobiological data, spurious overlaps collectively reduce the mean synaptic inputs to neurons, and that this mean reduction causes in turn an increase in storage capacity through a sparsening of network activity. Our paper demonstrates a link between a specific feature of experimentally inferred plasticity rules and network storage capacity.