Geometric organization of olfactory descriptor data in the Poincaré disk
Aniss Aiman Medbouhi, Farzaneh Taleb, Giovanni Luca Marchetti, Danica Kragic
cs.CE · cs.LG · stat.AP
Quick verdict
Paper maps odor descriptors into a Poincaré disk with radius~entropy, angle~categories; statistically careful but lacks a Euclidean MDS baseline.
Summary
The paper investigates whether a two-dimensional hyperbolic embedding gives an interpretable low-dimensional organization of odor descriptor data. The authors apply hyperbolic metric multidimensional scaling, optimized with Riemannian Adam, to two public datasets: 480 Sagar subject–odorant rating profiles (three subjects, 160 odorants, 15 continuous descriptors) and 4983 GoodScents–Leffingwell molecules with 138 binary expert labels. Embeddings preserve input pairwise distances moderately to well (distance Pearson 0.82 for Sagar, 0.69 for GSLF). In Sagar, rating profile entropy correlates strongly and negatively with hyperbolic radius (Pearson −0.77 ± 0.05 over 10 seeds; p < 0.001 under within-subject and odorant-block permutation tests), and the association survives removing intensity, pruning correlated descriptors, orthogonalization, subject-specific and averaged-rating analyses. Sweet, musky, fruity, and pleasantness show the strongest angular directional trends (mean R² 0.40–0.51). In GSLF, active label entropy increases with radius (r = 0.87), orthogonalized descriptor entropy decreases (r = −0.88), and related labels occupy coherent high-density regions shown via hyperbolic KDE. Main limitations: entropy is computed from the same vectors used to train the embedding, and no Euclidean MDS baseline tests whether the radial organization is hyperbolic-specific; only three subjects; the binary-descriptor region analysis is qualitative, as the authors acknowledge.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- In the Sagar union embedding, rating profile entropy is strongly negatively associated with hyperbolic radius (Pearson −0.77 ± 0.05 over 10 seeds, p < 0.001 under both restricted permutation schemes), with diffuse profiles near the center and concentrated profiles near the boundary.
- The radial organization is a profile-level property: only intensity among the 15 descriptors shows a radial association (0.42 ± 0.05), all others |r| < 0.2, and the entropy effect survives descriptor exclusion, pruning, and orthogonalization (Table 3).
- Sweet, musky, fruity, and pleasantness show the strongest angular directional trends in the tangent space (mean directional R² 0.40–0.51, Table 2), stable across subjects and averaged ratings.
- In GSLF, active label entropy increases with radius (r = 0.87 ± 0.03) while orthogonalized descriptor entropy decreases (r = −0.88 ± 0.01), and related binary descriptors occupy coherent, spatially separated high-density regions (Figures 6–7).
- Two-dimensional hyperbolic MDS is offered as a descriptive, interpretable framework — explicitly not a claim about the intrinsic dimensionality of olfaction or hyperbolic neural coding.
Strengths
Restricted permutation design matched to data structure
Within-subject and odorant-block permutation schemes (§2.4.2, following Winkler et al. 2014/2015) respect the 480-observation / 195-CID repeated-measures structure; B=1000 with the +1 correction (Phipson & Smyth) avoids zero p-values.
Unusually thorough robustness grid
Table 3 separates entropy recomputation on a fixed embedding from full retraining (excl.-intensity, pruned at |r|≤0.3, orthogonalized), keeping radial correlations in −0.68 to −0.75; Table 4 shows per-subject (−0.72/−0.86/−0.61) and averaged-rating (−0.82±0.08) replication.
Explicit, non-collapsing entropy semantics
Reproducibility and honest scoping
Weaknesses
No comparative embedding baseline
Entropy–embedding circularity
Arbitrary probability construction
Multiplicity uncorrected at descriptor level
GSLF result interpretation strain
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.09573, KTH) asks whether a two-dimensional Poincaré disk embedding of olfactory descriptor profiles is *interpretable*: does radial position encode global profile properties, and does angular position encode descriptor gradients and categories? Applying hyperbolic metric MDS to the Sagar dataset (480 subject–odorant rating profiles; 3 subjects; 195 unique CIDs; 15 common descriptors) and GoodScents–Leffingwell (4,983 molecules; 138 binary expert labels), the authors advance four claims: (1) Sagar rating profile entropy is strongly negatively associated with hyperbolic radius (Pearson −0.77 ± 0.05 over 10 seeds; p < 0.001 under both restricted permutation schemes); (2) this organization is profile-level, exceeding any single descriptor (only intensity shows a radial link, 0.42 ± 0.05; all others |r| < 0.2, Table S1); (3) sweet, musky, fruity and pleasantness carry the strongest angular directional trends (mean R² 0.40–0.51, Table 2); (4) in GSLF, active label entropy rises with radius (+0.87) while orthogonalized descriptor entropy falls (−0.88), and related labels form coherent high-density regions. §1 explicitly disclaims intrinsic-dimensionality and neural-coding claims — an unusually careful scope statement.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does the radius mean in the paper's Poincaré disk embedding of odors?
It is the hyperbolic distance from the disk's center. Empirically, in the Sagar human-rating data, diffuse rating profiles (high entropy) sit near the center and concentrated profiles near the boundary, with a correlation of about −0.77.
Which datasets does the study use, and are they public?
Two public datasets: the Sagar dataset via Pyrfume (480 subject–odorant ratings from 3 subjects on 160 odorants with 15 descriptors) and GoodScents–Leffingwell via OpenPOM (4,983 molecules with 138 binary expert labels).
Is there code to reproduce the hyperbolic odor embeddings?
Yes. The authors release code at github.com/anissmedbouhi/HyperSmell, and the paper specifies all hyperparameters: learning rate 0.1, 1000 training epochs, 10 random seeds, and 1000 permutations per test.
Does this paper prove that human olfactory space is hyperbolic?
No. The authors explicitly state the 2D embedding is a descriptive visualization and interpretation framework, not an estimate of olfaction's intrinsic dimensionality and not evidence that neural representations are two-dimensional or hyperbolic.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- PDF extraction flattens table layouts (rows of Table 1 and Table S1 run together in the raw text); every numeric value cited in this review remains attributable via the surrounding prose and captions.
- Equations are partially garbled by extraction, but all statistics quoted here were verified verbatim in the extracted text; the text ends with the Figure S1 caption and appears complete.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.