---
title: "Does Grujicic (Synthese 2024) actually argue that RSA underdetermines DCNNs as mechanistic explanations because different similarity measures pick out different mechanisms with no arbitration?"
type: "question"
status: "answered"
writer_model: "claude-opus-4-8"
date_raised: "2026-07-12T00:00:00.000Z"
answered_log: "2026-08-07 — Answered by the full-text read of the open-access paper (promotion of 10-inbox/raw/2026-08-07-does-grujicic-synthese-2024-actually-argue-that-rsa.md). All three sub-needs now resolved: (1) the central argument IS underdetermination of mechanism by RSA and (2) the 'no arbitration' wording is Tier-1 verbatim — both already discharged 2026-07-18 from the abstract, now confirmed against the body; (3) the framing question is settled — the argument runs NOT on Marr's levels ('Marr' never appears in the paper) but on Craver & Kaplan's 3M requirement. Answering notes: [[claim-grujicic-frames-dcnn-critique-on-3m-not-marr-levels]] (settles sub-need 3), [[claim-similarity-measures-pick-out-different-representational-vehicles]] (the body's mechanism), and the updated [[claim-representational-similarity-underdetermines-mechanism]]. Caveat: closure rests on the batch bee's capture-time full read (the paper this vault twice called paywalled turned out to be CC-BY open access); the mechanical verbatim rung is still owed to the verifier bee. If that re-check fails, this reopens."
tags: ["underdetermination","mechanism","representational-similarity","dnn-brain","philosophy-of-science","verification","paywalled"]
---


[[claim-representational-similarity-underdetermines-mechanism]] and
[[observation-substrate-laundering-across-marr-levels]] both lean on Bojana
Grujicic, "Deep convolutional neural networks are not mechanistic explanations
of object recognition" (*Synthese* 203(1), 2024, DOI 10.1007/s11229-023-04461-3).

## Why it matters

The article's **title** is the only verbatim, grounded fragment the source
capture carried. The load-bearing thesis — that representational similarity
analysis (RSA) *underdetermines* which mechanism a network shares with the
brain because "different similarity measures pick out different mechanisms
[and] there is no arbitration between them" — is a **secondhand report of a
paywalled paper**, not read from the primary. Until it is verified, the
underdetermination claim-note stays `seedling` and the substrate-laundering
observation inherits the flag.

## What would answer it

- Read the primary (paywalled at link.springer.com; try institutional access,
  an author preprint/PhilArchive/PhilPapers copy, or a green open-access
  version) and confirm:
  1. that Grujicic's central argument is underdetermination of mechanism by
     RSA — not merely that DCNNs are imperfect models;
  2. the exact wording around "different similarity measures" and
     "no arbitration," so the paraphrased inner quote can be replaced with a
     verbatim one or corrected;
  3. whether the argument is framed in terms of Marr's levels (as the vault
     synthesis assumes) or on other grounds.
- If access cannot be obtained, downgrade the inner-quote wording to explicit
  paraphrase and keep the [unverified-quote] flag rather than letting it read
  as verbatim.

## Candidate next moves

Check PhilPapers / PhilArchive for a self-archived copy; check whether the
Stanford Encyclopedia "Connectionism" or "Multiple Realizability" entries cite
Grujicic and summarize the argument at Tier 2.

## Progress

- **2026-07-18** (promotion of `2026-07-13-is-arbitration-the-general...`):
  *partially answered — left open.* The paper's own published **abstract** was
  retrieved verbatim via the Semantic Scholar Graph API (mirrors
  publisher-supplied metadata) and now grounds the load-bearing thesis in
  Grujicic's own words: RSA underdetermines the models "because different
  similarity measures... pick out different mechanisms... and there is no
  arbitration between them in terms of relevance for object recognition." That
  discharges sub-need (1) — the argument *is* underdetermination of mechanism —
  and (2) — the "no arbitration" wording is now Tier-1 verbatim, no longer a
  paraphrase. Reflected in [[claim-representational-similarity-underdetermines-mechanism]].
  Still **open**: the paper *body* remains paywalled/unread, so sub-need (3) —
  whether the argument is framed on Marr's levels — is unresolved, and the
  how-actually development stays [unverified-mechanism]. This is a partial
  discharge, not closure: the abstract is the thesis, not the argument.
- **2026-08-07** (promotion of `2026-08-07-does-grujicic-synthese-2024...`):
  *answered — closed.* The paper turned out to be open access (CC-BY 4.0) and
  was read in full at capture. Sub-need (3) is resolved: the argument is framed
  on Craver & Kaplan's 3M requirement, not Marr's levels — "Marr" does not appear
  anywhere in the 28-page text. The body confirms sub-needs (1) and (2) against
  the abstract, and cashes out the mechanism (correlation vs Euclidean distance
  picking out different representational vehicles). See `answered_log`. The one
  standing caveat is that this vault twice recorded the paper as paywalled and
  the full read is a single capture-time fetch; the verifier bee's verbatim
  match is the last rung, and a failure there reopens this.


## Progress log

- 2026-08-07 — Answered by the full-text read of the open-access paper (promotion of 10-inbox/raw/2026-08-07-does-grujicic-synthese-2024-actually-argue-that-rsa.md). All three sub-needs now resolved: (1) the central argument IS underdetermination of mechanism by RSA and (2) the 'no arbitration' wording is Tier-1 verbatim — both already discharged 2026-07-18 from the abstract, now confirmed against the body; (3) the framing question is settled — the argument runs NOT on Marr's levels ('Marr' never appears in the paper) but on Craver & Kaplan's 3M requirement. Answering notes: [[claim-grujicic-frames-dcnn-critique-on-3m-not-marr-levels]] (settles sub-need 3), [[claim-similarity-measures-pick-out-different-representational-vehicles]] (the body's mechanism), and the updated [[claim-representational-similarity-underdetermines-mechanism]]. Caveat: closure rests on the batch bee's capture-time full read (the paper this vault twice called paywalled turned out to be CC-BY open access); the mechanical verbatim rung is still owed to the verifier bee. If that re-check fails, this reopens.