Inferring the location of neurons within an artificial network from their activity

Inferring the connectivity of biological neural networks from neural activation data is an open problem. We propose that the analogous problem in artificial neural networks is more amenable to study and may illuminate the biological case. Here, we study the specific problem of assigning artificial neurons to locations in a network of known architecture, specifically the LeNet image classifier. We evaluate a supervised learning approach based on features derived from the eigenvectors of the activation correlation matrix. Experiments highlighted that for an image dataset to be effective for accurate localisation, it should fully activate the network and contain minimal confounding correlations. No single image dataset was found that resulted in perfect assignment, however perfect assignment was achieved using a concatenation of features from multiple image datasets. (c) 2022 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).

Automated summary

Inferring the connectivity of biological neural networks from neural activation data is challenging, but studying the analogous problem in artificial neural networks can provide insights into the biological case. This study focuses on assigning artificial neurons to locations in the LeNet image classifier. A supervised learning approach based on features derived from the activation correlation matrix is evaluated. The experiments suggest that an image dataset needs to fully activate the network and have minimal confounding correlations for accurate localization, and perfect assignment can be achieved by combining features from multiple image datasets.