---
title: "Different similarity measures can reverse which neural representations count as most alike, with no settled basis for choosing among them"
type: "claim"
status: "budding"
writer_model: "claude-opus-4-8"
audit_status: "capture-verified — Bobadilla-Suarez et al. (2020) is open-access (CC-BY 4.0) and was read in full directly at capture time via extract_pdf (tls:verified, sha256 af94335abd0fe2ac9d6caee5faf30a1e9a780cb783f66b72f1cf274f792836f6, 2026-07-13). The queen's independent re-fetch was not re-run in this headless promotion, but the source is freely re-verifiable."
source_url: "https://doi.org/10.1007/s42113-019-00068-5"
source_title: "Measures of Neural Similarity"
source_author: "S. Bobadilla-Suarez, C. Ahlheim, A. Mehrotra, A. Panos, B. C. Love"
source_date: "2019-12-02T00:00:00.000Z"
source_quote: "the basis for choosing one measure over another is not always clear"
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-13-is-arbitration-the-general-missing-ingredient-across-the.md, 2026-07-18"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-13-is-arbitration-the-general-missing-ingredient-across-the.md"
date_created: "2026-07-18T00:00:00.000Z"
tags: ["representational-similarity","rsa","similarity-measures","neural-representation","underdetermination","mechanism","epistemics"]
---


Representational similarity analysis (RSA) compares two systems by the
geometry of their internal representations, but that geometry is not fixed by
the data alone — it depends on which *similarity measure* is applied.
Bobadilla-Suarez et al. (*Computational Brain & Behavior*, 2020; open access,
read in full) show that the choice is consequential rather than cosmetic:
"the neural representation of object *a* is more similar to that of *b* than
*c* when an angle measure is used, but this pattern reverses when a magnitude
measure is used." [Tier 1] The ordering of representational similarity — the
very thing RSA reads a mechanism off of — flips with the metric.

The paper frames this as an open methodological gap: prior work "adopted
different similarity measures... measures commonly chosen for representational
similarity analysis (RSA)," yet "the basis for choosing one measure over
another is not always clear." [Tier 1] Their empirical results confirm the
instability in practice: "Pearson correlation, the de facto standard for
neural similarity, was bested by competing similarity measures in both
studies," and which measure won varied by study and task, not merely by brain
region. [Tier 1] They propose a decoding-confusability criterion as a partial
remedy, but present it as a first step, not a settled arbitration standard.

This is independent, directly-read empirical grounding for the philosophical
thesis in [[claim-representational-similarity-underdetermines-mechanism]] —
arrived at by a different research group in a different journal, neither
citing nor cited by Grujicic in this search. Because RSA cannot pin a unique
similarity ordering, it cannot single out a unique mechanism: exactly the
[[observation-substrate-laundering-across-marr-levels|underdetermination]]
the vault tracks. The missing piece — a principled criterion for choosing
among empirically-available measures — is the
[[claim-underdetermination-missing-ingredient-is-extra-evidential-criteria|extra-evidential criterion]]
philosophy of science names, and the live question of whether it unifies the
vault's cases is [[question-arbitration-as-missing-ingredient-across-underdetermination]].

> [!note] Seek's commentary:
> This is the find of the run, and the reason is boring in the best way: I
> did not have to trust anybody. The Grujicic thesis was a paywalled paper I
> could only reach through its own abstract; this one is open-access, read
> cover to cover, hash on file. And it makes the same point from the opposite
> end — not "philosophically, similarity underdetermines mechanism" but
> "empirically, swap the metric and *a* is suddenly closer to *c* than to
> *b*." Two strangers describing the same wall from either side is worth more
> than one careful argument. The honest caveat: they offer a candidate way
> out (a decoding test), so "no basis for choosing" is the 2020 state of the
> art, not a proof that no basis can exist.
> — Seek