Associative Recall in Non-Randomly Diluted Neuronal Networks

TL;DR

This paper investigates how the spatial distribution and network architecture of diluted neuronal networks affect their ability to perform associative recall, finding that more uniform cell distributions improve performance.

Contribution

It introduces biologically relevant configurations of neuronal networks, including scale-free and geometrically diluted architectures, and analyzes their impact on associative recall.

Findings

Network performance improves with spatial uniformity of cell distribution.

Scale-free and geometrically diluted architectures are effective for associative recall.

Spatially denser regions tend to enhance network connectivity and recall ability.

Abstract

The potential for associative recall of diluted neuronal networks is investigated with respect to several biologically relevant configurations, more specifically the position of the cells along the input space and the spatial distribution of their connections. First we put the asymmetric Hopfield model onto a scale-free Barabasi-Albert network. Then, a geometrical diluted architecture, which maps from L-bit input patterns into $N$-neurons networks, with R=N/L<1 (we adopt R=0.1, 0.2 and 0.3), is considered. The distribution of the connections between cells along the one-dimensional input space follows a normal distribution centered at each cell, in the sense that cells that are closer to each other have increased probability to interconnect. The models also explicitly consider the placement of the neuronal cells along the input space in such a way that denser regions of that space tend to become denser, therefore implementing a special case of the Barabasi-Albert connecting scheme. The obtained results indicate that, for the case of the considered stimuli and noise, the network performance increases with the spatial uniformity of cell distribution.