In Grujicic's account, correlation and Euclidean distance pick out different representational vehicles, so RSA's choice of similarity measure selects which mechanism-type is being compared

representational-similarityrsasimilarity-measuresmechanismrepresentational-vehicledcnn-brainphilosophy-of-scienceepistemics

The load-bearing thesis that RSA underdetermines shared mechanism is not a slogan resting on the paper's abstract; Grujičić (Synthese 203:30, 2024) cashes it out mechanistically. She distinguishes two rival accounts of what counts as the representational vehicle a mechanism is typed by — a Tuning Functions account (individual neuron/node response profiles) and a Neural Manifolds account (population-level response geometry) — and argues, in the section that sets out both accounts, that correlation (an angle-based measure) and Euclidean distance (a magnitude-based measure) "pick out different representational vehicles, according to the Tuning Functions and the Neural Manifolds accounts" (Sect. 5, p.16). [Tier 1, verbatim] She restates the conclusion in the paper's assessment section: the similarity measures serving as RSA's mapping function "pick out different types of mechanisms on the level of representational vehicles" (Sect. 7, p.20). [Tier 1, verbatim] The choice of similarity measure is therefore not a downstream analysis detail but the step that selects which mechanism-type is being compared across a network and the brain.

The worked example is a re-analysis of Ramírez et al. (2014) on face-orientation coding in the human fusiform face area: "When the authors used correlation to quantify the representational geometries, the unimodal tuning hypothesis was favoured... However, when Euclidean distance was used, the two hypotheses were indistinguishable." (Sect. 4.3, p.15) [Tier 1, verbatim] The same data support different theoretical conclusions depending on the metric. Grujičić's conclusion: "it is underdetermined which of the mechanism types picked out by similarity measures are relevant for object recognition." (Sect. 7, p.21) [Tier 1, verbatim]

This is the primary-side mechanism behind the empirical instability that Bobadilla-Suarez et al. (2020) reach from the data side — an angle-vs-magnitude taxonomy Grujičić cites explicitly (Sect. 4.1) as the source of the distinction. Two independent lines converging on one wall: swap the measure and the mechanism-type changes, with nothing in RSA to arbitrate which measure is relevant. It is the same cross-level underdetermination the vault tracks.