Does Grujicic (Synthese 2024) actually argue that RSA underdetermines DCNNs as mechanistic explanations because different similarity measures pick out different mechanisms with no arbitration?

underdeterminationmechanismrepresentational-similarityrsadcnn-brainphilosophy-of-sciencemarr-levelsepistemicsverification

This capture directly answers question-verify-grujicic-2024-rsa-underdetermination-dcnn, the open verification question routed by claim-representational-similarity-underdetermines-mechanism. Prior to this run, the vault's grounding for the paper's load-bearing thesis rested on the published abstract only (retrieved via the Semantic Scholar Graph API), with the paper body — the argument's actual development, and whether it invokes Marr's levels — unread and flagged [unverified-mechanism]. The article turned out to be open access; it was fetched in full (28 pages, tls: verified) via extract_pdf from Springer's own content route (link.springer.com/content/pdf/..., the paper's original publisher venue, not a scraper mirror) and read cover to cover. All quotes below were checked against the extracted text with quote_check and returned grounded.

Claim: Yes — Grujicic's own paper states the underdetermination-by-arbitration thesis exactly as the vault's prior capture attributed to her, verbatim and as the paper's central argument, not a paraphrase

The paper's abstract states the thesis in the author's own words: "I focus on one frequent method of their comparison—representational similarity analysis, and I argue, first, that it underdetermines these models as how-actually mechanistic explanations. This happens because different similarity measures in this framework pick out different mechanisms across DCNNs and the brain in order to correspond them, and there is no arbitration between them in terms of relevance for object recognition." [Tier 1, verbatim] Section 7, titled "DCNNs are not how-actually mechanism schemata of object recognition," restates the conclusion directly: "If there is no arbitration between them in terms of relevance for object recognition, it is clear that current DCNNs are not how-actually mechanism schemas of object recognition." [Tier 1, verbatim] This closes sub-needs (1) and (2) of the open question at Tier 1 rather than Tier 1-via-abstract-only: the thesis is confirmed to be the paper's actual argument, not an abstract that oversells a narrower body argument.

Claim: The "different measures pick out different mechanisms" move is cashed out mechanistically — correlation and Euclidean distance pick out non-identical representational vehicles under two rival accounts of what a vehicle is

This is the technical-mechanism content behind the thesis, which the abstract alone could not verify. Grujicic distinguishes two candidate 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 correlation (an angle-based measure) and Euclidean distance (a magnitude-based measure) "pick out different types of mechanisms on the level of representational vehicles" under both accounts (Sect. 5, p.16). [Tier 1, verbatim] She grounds this empirically in a re-analysis of Ramirez et al. (2014), a study of face-orientation coding in human fusiform face area (FFA): "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] Her conclusion from this case study: "Given that the relevance of similarity measures used as a part of the RSA framework is not assessed, 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 mechanism, not just the abstract's slogan, and it directly cites and builds on the correlation-vs-Euclidean (angle-based vs. magnitude-based) taxonomy from Bobadilla-Suarez et al. 2020, which Grujicic cites explicitly and repeatedly (Sect. 4.1) as the source of that taxonomy — confirming, from the primary side this time, that the two vault notes are independently converging on the same empirical mechanism rather than merely rhyming.

Claim: The argument is NOT framed on Marr's levels of analysis, contrary to the vault synthesis's prior assumption — it is framed via Craver & Kaplan's model-to-mechanism-mapping (3M) requirement and the how-actually/how-possibly mechanistic-schema distinction

This resolves sub-need (3) of the open question, which asked "whether the argument is framed in terms of Marr's levels (as the vault synthesis assumes) or on other grounds." Having read the full 28-page text, the word "Marr" does not appear anywhere in the paper. [Tier 1, absence confirmed by direct full read] Instead, Grujicic's explicit theoretical apparatus is the mechanistic-explanation literature: the "model-to-mechanism mapping" (3M) requirement from Kaplan & Craver (2011) and Craver & Kaplan (2020) — "A model of object recognition has mechanistic explanatory force if it has variables that map onto the representations, activities, and organisational properties of the brain mechanism... This is the model-to-mechanism mapping (3M) requirement for mechanistic explanations" (Sect. 3.1, p.7) [Tier 1, verbatim] — together with the "mechanism schema" concept from Machamer, Darden & Craver (2000) and Darden (2002), and the how-actually/how-possibly distinction for mechanistic explanatory models (Craver 2007; Brainard 2020) that structures the paper's two main sections (7 and 8). The vault's prior synthesis note's assumption that the argument runs on Marr's levels should be corrected going forward: it runs on the Craver/Kaplan mechanistic-explanation framework instead. This discharges the last open flag on claim-representational-similarity-underdetermines-mechanism and on question-verify-grujicic-2024-rsa-underdetermination-dcnn.

Further leads

• Grujicic's paper makes a second, independent argument beyond the RSA-arbitration thesis (Sect. 8): DCNNs also fail as how-possibly mechanistic explanations because their architectural idealization is too under-constrained to bound the space of candidate mechanisms — contrasted with Riesenhuber & Poggio's (2000) HMAX model as a genuine how-possibly explanation. Outside this capture's scope (which is the RSA-arbitration question specifically) but a distinct claim worth its own note if the vault wants full coverage of the paper's argument.
• Bojana Grujičić and Phyllis Illari, "Using deep neural networks and similarity metrics to predict and control brain responses," forthcoming in The Routledge Handbook of Causality and Causal Methods (Illari & Russo, eds.) — self-archived preprint at philsci-archive.pitt.edu/id/eprint/22797 (not read this session; the PhilSci-Archive page returned "Request Rejected" to both archive_page and WebFetch, possibly a WAF/bot-block rather than adversarial content — worth a retry from a different route). Likely extends the same similarity-measure analysis toward prediction/control framings rather than explanation.
• Grujicic explicitly engages, and argues against, two prior philosophical defenses of DCNNs-as-mechanistic-explanations: Cao & Yamins (2021a, 2021b, arXiv preprints later published as "Explanatory models in neuroscience" Parts 1–2) and Buckner (2018, Synthese 195(12):5339–5372, "Empiricism without magic"). Neither read directly this session; both are the direct targets the paper is written to rebut.
• Roskies (2021, Synthese 199(3-4):5917-5935) and Kieval (2022, Synthese 200(3), "Mapping representational mechanisms with deep neural networks") are named by Grujicic (Sect. 6) as making the inference she calls fallacious — that matching representational geometries licenses matching representational content. A second, related but distinct underdetermination claim (geometry-match does not fix content, separate from geometry-match not fixing mechanism-type) that could be its own atomic note.

Entity candidates

• Nikolaus Kriegeskorte — person — originator of representational similarity analysis itself (Kriegeskorte, Mur & Bandettini 2008), the method the entire paper interrogates; the foundational figure Grujicic's critique is ultimately aimed at, not merely a citation. Flagging first per the ancestry-figure blind spot.
• Machamer, Darden & Craver (Peter Machamer, Lindley Darden, Carl F. Craver) — persons/concept — authors of "Thinking about mechanisms" (Philosophy of Science, 2000), the foundational mechanistic-explanation framework (mechanism schemas, entities/activities) that supplies Grujicic's theoretical apparatus in place of Marr's levels.
• David Kaplan and Carl Craver — persons — source of the "model-to-mechanism mapping" (3M) requirement (Kaplan & Craver 2011; Craver & Kaplan 2020) that is the specific mechanistic-adequacy test Grujicic applies to DCNNs throughout the paper.
• Chris Buckner — person — philosopher whose 2018 Synthese paper argues DCNNs and the visual cortex instantiate a shared mechanism; one of the two direct philosophical targets Grujicic's paper is written to rebut.
• Rosa Cao and Daniel Yamins — persons — authors of "Explanatory models in neuroscience" (Parts 1-2), who argue DCNNs satisfy the 3M requirement and are already mechanistic explanations; the other direct philosophical target of this paper.
• Bojana Grujičić — person — author of the paper this capture verifies; Max Planck School of Cognition / Humboldt-Universität zu Berlin / UCL Science and Technology Studies at time of publication.
• Fernando M. Ramírez — person — author of the empirical FFA case study (Ramirez et al. 2014, Journal of Neuroscience) that Grujicic uses as her central worked example of correlation and Euclidean distance yielding contradictory theoretical conclusions from the same data.