Obstacles to inferring mechanistic similarity using Representational Similarity Analysis

Abstract

A core challenge in neuroscience is to assess whether diverse systems represent the world similarly. Representational Similarity Analysis (RSA) is an innovative approach to address this problem and has become increasingly popular across disciplines from machine learning to computational neuroscience. Despite these successes, RSA regularly uncovers difficult-to-reconcile and contradictory findings. Here we demonstrate the pitfalls of using RSA to infer representational similarity and explain how contradictory findings arise and support false inferences when left unchecked. By comparing neural representations in primate, human and computational models, we reveal two problematic phenomena that are ubiquitous in current research: a 'mimic' effect, where confounds in stimuli can lead to high RSA scores between provably dissimilar systems, and a 'modulation effect', where RSA-scores become dependent on stimuli used for testing. Since our results bear on existing findings and inferences, we provide recommendations to avoid these pitfalls and sketch a way forward.

Deep problems with neural network models of human vision

“…If something walks like a duck and quacks like a duck, isn't it in all likelihood a duck? In fact, DNNs often make their predictions in unexpected ways, exploiting "short-cuts" that humans never rely on (e.g., Geirhos et al, 2018;Malhotra, Evans, and Bowers, 2020;Malhotra, Dujmović, & Bowers, 2022;Rosenfeld, Zemel, Tsotsos, 2018). For example, Malhotra et al (2020) systematically inserted single pixels (or clouds of pixels) into photographs in ways that correlated with image category so that the images could be classified based on the photographic subjects themselves or the pixels.…”

“…The critical issue for present purposes, however, is whether models that classify images based on short-cuts also perform well on prediction-based experiments. Dujmović et al (2022) explored this question using RSA which compares the distances between activations in one system to the distances between corresponding activations in the second system (see Figure 3). To compute RSA, two different systems (e.g., DNNs and brains) are presented the same set of images and the distance between the representations for all pairs of images is computed.…”

“…This representational geometry is expressed as a representational dissimilarity matrix (RDM) for each system. Finally, an RSA score is determined by computing the correlation between the two RDMs (image taken from Dujmović et al, 2022).…”

“…To assess whether large RSAs can be obtained between two very different systems, Dujmović et al (2022) carried out a series of simulations that computed RSAs between two DNNs or between DNNs and single-cell recordings from macaque IT when the two systems classified objects in qualitatively different ways. For example, when comparing DNNs to macaque IT, the authors trained a DNN to classify photographs taken from Majaj, Hong, Solomon, and DiCarlo (2015) that contained a pixel patch confound (call it DNN-pixel) as well as unperturbed photos (DNNstandard), similar to the Malhotra et al (2020) setup described above.…”

On Logical Inference over Brains, Behaviour, and Artificial Neural Networks

“…A series of mainstream methodological techniques used in CCN that were developed originally by another subfield of cognitive science, specifically mathematical psychology (see Navarro, 2021;Shepard & Chipman, 1970), have shown that computing correlations over correlations can provide useful insights in terms of the structure and relationships between and within stimulus representations and between and within different organisms and models (cf., Dujmović et al, 2022). For example, we "correlate a brain region's RDM [representational dissimilarity matrix; a second-order isomorphism of internal representations] with an RDM based on one or multiple stimulus parameters (or with an RDM predicted by a computational model), [to obtain] the correlation between the two RDMs."…”