Interpreting weight maps in terms of cognitive or clinical neuroscience: nonsense?

TL;DR

This paper investigates the reliability of weight maps in neuroimaging machine learning models, revealing that false positives are unlikely under certain conditions and questioning their use as direct neural signal proxies.

Contribution

It provides an empirical analysis of how signal-to-noise ratio and sparsity influence the interpretation of weight maps in ECoG data, challenging assumptions about their neural relevance.

Findings

False positives in weight maps are not common across all conditions.

High weight features are unlikely to be false positives in most cases.

Signal-to-noise ratio and sparsity affect the similarity between neural signals and weight maps.

Abstract

Since machine learning models have been applied to neuroimaging data, researchers have drawn conclusions from the derived weight maps. In particular, weight maps of classifiers between two conditions are often described as a proxy for the underlying signal differences between the conditions. Recent studies have however suggested that such weight maps could not reliably recover the source of the neural signals and even led to false positives (FP). In this work, we used semi-simulated data from ElectroCorticoGraphy (ECoG) to investigate how the signal-to-noise ratio and sparsity of the neural signal affect the similarity between signal and weights. We show that not all cases produce FP and that it is unlikely for FP features to have a high weight in most cases.