Pith. sign in

REVIEW 3 major objections 6 minor 18 references

CuBAS: Information Geometric Curvature-Based Adaptive Sampling for Supervised Classification

T0 review · 3 major / 6 minor · reviewed 2026-07-12 · grok-4.5

Pith's one-line read Local curvature of a Potts statistical manifold on a labeled graph identifies the most informative samples for classification, beating random and uncertainty sampling across 60+ datasets.

desk verdict Clean Potts-Fisher curvature scores for mixed high/low sampling, with heavy k-NN wins on 60+ datasets, but the multi-architecture claim and thin baselines leave the general informativeness story incomplete. read the letter →

arxiv 2607.03145 v1 pith:NZSAZFVW submitted 2026-07-03 cs.LG cs.AIcs.CVcs.ITmath.ITstat.ML

classification cs.LGcs.AIcs.CVcs.ITmath.ITstat.ML
keywords adaptivesamplinginformationgeometryPottsmodelFishercurvaturek-NNgraphsupervisedclassificationshapeoperator
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Most ways of shrinking a labeled training set ignore the geometry of how labels interact. This paper claims that if you view a labeled dataset as a statistical manifold induced by a q-state Potts model on a k-nearest-neighbor graph, a simple local curvature score—the ratio of second- to first-order observed Fisher information—separates smooth cluster interiors from decision-boundary neighborhoods. Selecting a mixture of low-curvature prototypes and high-curvature boundary points produces compact training subsets that raise balanced accuracy under a k-NN classifier more than random sampling or entropy-based uncertainty sampling. The gains hold across more than 60 OpenML datasets spanning tabular, image, genomic and spectroscopic domains, and across labeling budgets from 10% to 80%, with the largest relative lift in the scarcest-label regime. The procedure is closed-form once the graph and a single inverse-temperature parameter are known, scales linearly with the number of graph edges, and needs no auxiliary classifier. A sympathetic reader cares because the result turns sample selection into a geometric inference problem rather than a heuristic or model-dependent score, offering a practical route to data-efficient supervised learning when labels are costly.

What carries the argument

Local shape operator Si of the one-dimensional Potts statistical manifold: the closed-form ratio of node-wise second-order to first-order observed Fisher information, which measures how rapidly the local energy landscape bends with the inverse-temperature parameter and thereby ranks geometric complexity around each vertex.

What would settle it

On the same 60+ OpenML suite and protocol, replace the k-NN evaluator with a logistic or tree-based classifier (or construct the graph in a non-Euclidean embedding) and check whether the balanced-accuracy advantage of CuBAS over entropy and random sampling disappears or reverses.

Watch

Extended reading notes

Core claim

The local scalar shape operator Si = −Ψi/(Φi + λ) computed from the first- and second-order observed Fisher information of an isotropic q-state Potts Markov random field on a labeled k-NN graph partitions samples into complementary low- and high-curvature regimes; mixed subsets drawn from both regimes are more informative for classification than randomly or uncertainty-selected subsets of the same size.

Load-bearing premise

That Euclidean k-nearest-neighbor geometry both defines the Potts field and fairly evaluates sample informativeness for classifiers in general, even though the reported experiments use only a k-NN classifier.

Editorial extensions

If this is right

  • Compact, geometry-aware training subsets can replace random or uncertainty sampling without needing an auxiliary model.
  • High-curvature samples concentrate near class interfaces and can be used as a filter for boundary-aware learning.
  • Low-curvature samples act as stable prototypes, enabling redundancy reduction inside homogeneous clusters.
  • The same curvature field can be recomputed after each new label batch, supporting iterative active selection.
  • Linear cost in graph edges makes the method usable on moderately large labeled pools where repeated classifier training is expensive.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the curvature signal remains informative in learned deep embeddings, CuBAS could serve as a coreset or warm-up set constructor for neural-network training.
  • The same Fisher-ratio curvature may transfer to semi-supervised or multi-modal graphs once edges encode heterogeneous proximity.
  • When class overlap is extreme and n is tiny, regularizing the curvature field over the graph Laplacian may restore contrast between the two regimes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes CuBAS, a geometry-aware sampling method for supervised classification. A labeled dataset is represented as a k-NN graph and modeled by an isotropic q-state Potts MRF; local first- and second-order observed Fisher information (Φi, Ψi) yield a scalar shape-operator score Si=−Ψi/(Φi+λ) at each vertex. Normalized curvature is adaptively thresholded (Otsu + Ashman D, with Tukey-IQR fallback) into low- and high-curvature regimes, and a mixed subset is drawn from both. The claim is that these subsets are more informative than random or entropy-based selection. Empirically, under a fixed k-NN (k=5) classifier, CuBAS reports higher average balanced accuracy on 60+ OpenML datasets across 15 training fractions (10–80%), with strong gains at the 10% budget (Wilcoxon p<10−10).

Significance. If the multi-architecture and architecture-agnostic informativeness claims hold, CuBAS would be a useful, model-free alternative to uncertainty sampling and classical prototype selection, with closed-form Potts Fisher scores, linear cost in graph edges after k-NN construction, adaptive thresholding, and public code. The empirical protocol (many datasets × budgets × 100 splits) is unusually thorough for instance-selection work and is a genuine strength under the reported evaluator. The information-geometric framing (1D shape operator from observed Fisher information) is clear and reproducible. The main open question is whether the gains survive when the evaluator does not share the Euclidean k-NN geometry used to define the Potts field.

major comments (3)
  1. Abstract and §1 claim improvements “across … classifier architectures” and “multiple classification algorithms,” but §4 and Tables 5–7 evaluate exclusively with a k-NN classifier (k=5). No other model (linear, tree, SVM, neural net) is trained on the same subsets. The multi-architecture claim is therefore unsupported as written and must be removed or backed by additional experiments.
  2. §4 protocol / Alg. 1 / §3.1–3.2: the evaluation metric reuses the same Euclidean k-NN neighborhoods that define Ui, Φi, Ψi, and Si. High-curvature nodes are precisely those with heterogeneous label neighborhoods under that metric, so Tables 5–7 and the Wilcoxon result at 10% budget may partly reflect sampler–evaluator coupling rather than architecture-agnostic informativeness. At least one non-neighbor-based classifier on the same CuBAS subsets is needed to decouple geometry from evaluation.
  3. §4 baselines are only random and entropy-based uncertainty sampling. The introduction positions CuBAS against prototype/instance selection and graph-based methods (Wilson, Garcia et al., etc.), yet none of those standard instance-reduction baselines appear in Tables 5–7. Without them, the claim of a “principled, scalable alternative to heuristic sampling” is only partially tested.
minor comments (6)
  1. Abstract inconsistency: the arXiv-style abstract says “more than 60” datasets; the body abstract says “30.” Tables 2–4 list ~68. Align the counts.
  2. §2.2: typo “pseudo-likeilhood”; elsewhere “Looking carefully to the pseudo-likeilhood equation.”
  3. Free parameters (graph k=15, λ=10−4, Dmin=2.6, μ2 floor 0.1, Otsu bins B, evaluation k=5) are fixed without sensitivity analysis. A short ablation on k and Dmin would strengthen §3.3–3.4.
  4. Figure 2 caption mentions four datasets; the surrounding text says “two representative datasets.”
  5. §3.1–3.2: Φi and Ψi are observed (empirical) Fisher terms, not expectations; the text sometimes calls them Fisher information without the “observed” qualifier—worth standardizing.
  6. Code and OpenML data availability are stated clearly; keep the GitHub link and note the exact commit or release used for the tables.

Circularity Check

0 steps flagged · score 0.0 of 10

No definitional or fit-as-prediction circularity: Potts Fisher curvature is an independent score, and accuracy gains are external empirical tests, not forced by construction.

full rationale

The load-bearing chain is: labeled k-NN graph → isotropic Potts local conditionals (Eq. 1) → MPL estimate of β (Eqs. 2–5) → node-wise first- and second-order observed Fisher Φi, Ψi (Eqs. 16–24) → scalar shape operator Si = −Ψi/(Φi+λ) (Eq. 31) → adaptive partition of H/L and mixed sampling (Alg. 1–2) → balanced accuracy vs random/entropy on held-out OpenML splits (Tables 5–7). None of these steps reduces to its own input by algebra or by re-labeling a fit as a prediction. Si is a closed-form functional of the Potts sufficient statistics Ui and β; it does not encode the downstream k-NN balanced accuracy, so the reported gains (including Wilcoxon p<10−10 at 10% budget) are independent empirical outcomes, not tautologies. Self-citation of Levada et al. (2010) only supplies the standard Potts conditional form, not a uniqueness theorem that forces the sampling claim. Concerns that the evaluator reuses the same Euclidean k-NN geometry as the sampler, or that the abstract overclaims multi-architecture support while only k=5 NN is reported, are experimental-coupling / overclaim issues, not circularity under the required Eq.-X≡Eq.-Y or fitted-input-as-prediction standard. The derivation is self-contained against external benchmarks.

Assumptions & free parameters 6 free parameters · 5 assumptions · 2 invented entities

The method rests on standard MRF/info-geometry tools plus modeling choices that turn local Potts Fisher ratios into a sampling criterion. Free parameters are graph and threshold knobs fixed by the authors; axioms are the usual regularity for Fisher information plus the domain claim that Potts curvature tracks classification informativeness; the main invented construct is the operational local shape score and the CuBAS partition itself.

free parameters (6)
  • k (k-NN graph degree) = 15
    Fixed at 15 in experiments; controls neighborhood size for Ui and thus all curvature scores.
  • λ (shape-operator regularizer) = 1e-4
    Added in Si = −Ψi/(Φi+λ) for numerical stability; set to 1e-4 without sensitivity study.
  • D_min (Ashman bimodality cutoff) = 2.6
    Conservative threshold 2.6 (literature often uses 2) gates Otsu vs IQR path for T*.
  • μ2 floor for high-curvature GMM mode = 0.1
    Rejects Otsu if high mode mean < 0.1; ad hoc guard against spurious bimodality.
  • histogram bins B for Otsu = 256
    Default 256 for discrete search of T_Otsu.
  • evaluation k-NN neighbors = 5
    Classifier k=5 fixed for all reported balanced accuracies.
assumptions (5)
  • domain assumption Labeled data on a k-NN graph is well modeled by an isotropic q-state Potts MRF with a single inverse temperature β.
    §2.1–2.2; all local energies Ui and Fisher scores assume this homogeneous pairwise interaction.
  • domain assumption Maximum pseudo-likelihood yields a usable β for observed Fisher information on finite graphs.
    §2.2 cites consistency results; finite-sample quality is assumed throughout experiments.
  • ad hoc to paper For a 1D statistical manifold, the scalar shape operator −II/I (here −Ψi/Φi) measures local geometric complexity relevant to sample informativeness.
    §2.3.3 and Def. 4 / Eq. (31); classical DG object is repurposed as a ranking score for classification without a risk bound.
  • ad hoc to paper Mixing low- and high-curvature nodes at the same fraction p produces compact yet maximally informative training sets.
    Algorithm 1 steps 12–16; design choice, not derived from a sample-complexity theorem.
  • standard math Standard regularity conditions allow interchange of derivative and expectation for Fisher information of the Potts conditionals.
    §3 Definitions 2–3 and subsequent derivations.
invented entities (2)
  • Local scalar Potts shape operator Si(β) as per-node sampling score
    purpose: Rank vertices by geometric complexity for adaptive train/test partition.
    Eq. (31); operational definition combining observed Φi, Ψi; independent evidence is only the downstream accuracy tables, not an external physical measurement.
  • CuBAS adaptive threshold cascade (Otsu + Ashman D + Tukey IQR)
    purpose: Choose T* without fixed quantiles across heterogeneous curvature histograms.
    Algorithm 2; engineered decision rule with hand-set Dmin and μ2 floor.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CuBAS: Information Geometric Curvature-Based Adaptive Sampling for Supervised Classification." pith.science (2026). https://pith.science/paper/NZSAZFVW

@misc{pith2026260703145,
  author       = {Pith},
  title        = {Pith review of: CuBAS: Information Geometric Curvature-Based Adaptive Sampling for Supervised Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NZSAZFVW}},
  note         = {Machine review of arXiv:2607.03145}
}
read the original abstract

The informativeness of a training set is as consequential as its size, yet most sampling strategies remain agnostic to the intrinsic geometry of the data distribution. We introduce CuBAS (Curvature-Based Adaptive Sampling), an information-geometric framework for adaptive data selection in supervised classification, grounded in the q-state Potts Markov random field (MRF) model. The central insight is that a labeled dataset can be viewed as a statistical manifold, on which local curvature, estimated via the ratio of second to first-order observed Fisher information, faithfully encodes the geometric complexity of the data distribution. We construct a k-nearest-neighbor graph over the labeled data and derive a closed-form curvature score at each vertex from the Potts sufficient statistics. This curvature signal partitions the graph into two complementary regimes: low-curvature regions, corresponding to smooth, homogeneous clusters, and high-curvature regions, concentrated around decision boundaries that are disproportionately informative for classification. By selecting nodes from both regimes, CuBAS constructs compact yet maximally informative training subsets. Empirical evaluation across more than 60 benchmark datasets demonstrates consistent and statistically significant improvements over random sampling and uncertainty-based baselines, across a wide range of labeling budgets and classifier architectures. CuBAS is computationally efficient (linear in the number of k-NN graph edges), theoretically grounded in the differential geometry of statistical manifolds, and interpretable in terms of the local shape operator of the data manifold.

Figures

Figures reproduced from arXiv: 2607.03145 by the authors.

Figure 1
Figure 1. Geometric interpretation of the shape operator. As one moves along a tangent direction [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Visualization of high-curvature nodes (H-nodes) identified by CuBAS in the digits, mfeat_karhunen, semeion and wine datasets. High-curvature samples are highlighted in black and concentrate around class boundaries, where inter-class interactions are most prominent. in the local label configuration. These regions typically correspond to class interfaces and decision boundaries, and therefore capture the most discrimi… view at source ↗
Figure 3
Figure 3. Average balanced accuracies achieved by the k-NN classifier under entropy-based uncertainty sampling, and [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Average balanced accuracies obtained by random sampling, entropy-based sampling and the proposed CuBAS [PITH_FULL_IMAGE:figures/full_fig_p024_4.png]
Figure 5
Figure 5. Figure 5: Average balanced accuracies obtained by random sampling, entropy-based sampling and the proposed CuBAS [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 6
Figure 6. Figure 6: Average balanced accuracies obtained by random sampling, entropy-based sampling and the proposed CuBAS [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 4 canonical work pages

  1. [1]

    Salvador Garcia, Joaquin Derrac, Jose Cano, and Francisco Herrera

    doi: 10.1109/TSMC.1972.4309137. Salvador Garcia, Joaquin Derrac, Jose Cano, and Francisco Herrera. Prototype selection for nearest neighbor classifi- cation: Taxonomy and empirical study.IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(3): 417–435,

  2. [2]

    Xiaojin Zhu, Zoubin Ghahramani, and John Lafferty

    doi: 10.1109/TPAMI.2011.142. Xiaojin Zhu, Zoubin Ghahramani, and John Lafferty. Semi-supervised learning using gaussian fields and harmonic functions. InProceedings of the Twentieth International Conference on International Conference on Machine Learning, ICML’03, page 912–919. AAAI Press,

  3. [3]

    27 PREPRINT- JULY7, 2026 Mikhail Belkin, Partha Niyogi, and Vikas Sindhwani

    ISBN 1577351894. 27 PREPRINT- JULY7, 2026 Mikhail Belkin, Partha Niyogi, and Vikas Sindhwani. Manifold regularization: A geometric framework for learning from labeled and unlabeled examples.Journal of Machine Learning Research, 7(85):2399–2434,

  4. [4]

    doi: 10.1109/MSP.2017. 2693418. Mathilde Papillon, Sophia Sanborn, Johan Mathe, Louisa Cornelis, Abby Bertics, Domas Buracas, Hansen J. Lillemark, Christian Shewmake, Fatih Dinc, Xavier Pennec, and Nina Miolane. Beyond Euclid: An illustrated guide to modern machine learning with geometric, topological, and algebraic structures.Machine Learning: Science an...

  5. [5]

    Melanie Weber

    doi: 10.1088/2632-2153/adf375. Melanie Weber. Geometric machine learning.AI Magazine, 46(1):e12210,

  6. [6]

    doi: https://doi.org/10.1002/aaai. 12210. Shun-ichi Amari.Information Geometry and Its Applications. Springer,

  7. [7]

    doi: 10.3390/e22101100

    ISSN 1099-4300. doi: 10.3390/e22101100. URLhttps://www.mdpi.com/1099-4300/22/10/1100. Simon Tong and Daphne Koller. Support vector machine active learning with applications to text classification.J. Mach. Learn. Res., 2:45–66, March

  8. [8]

    doi: 10.1162/153244302760185243

    ISSN 1532-4435. doi: 10.1162/153244302760185243. Burr Settles. Active learning literature survey.University of Wisconsin–Madison Technical Report,

Show all 18 references
  1. [9]

    doi: 10.1145/1553374.1553380

    Association for Computing Machinery. doi: 10.1145/1553374.1553380. H. Liu and H. Motoda. On issues of instance selection.Data Mining and Knowledge Discovery, 6:115–130,

  2. [10]

    URLhttps://doi.org/10.1023/A:1014056429969. C.C. Aggarwal.Data Classification: Algorithms and Applications. Chapman and Hall/CRC,

  3. [11]

    doi: https: //doi.org/10.1201/b17320. J. Arturo Olvera-López, J. Ariel Carrasco-Ochoa, J. Francisco Martínez-Trinidad, and Josef Kittler. A review of instance selection methods.Artificial Intelligence Review, 34:133–143,

  4. [12]

    doi: https://doi.org/10.1016/j

    ISSN 0031-3203. doi: https://doi.org/10.1016/j. patcog.2019.07.004. Dengyong Zhou, Olivier Bousquet, Thomas Navin Lal, Jason Weston, and Bernhard Schölkopf. Learning with local and global consistency. InProceedings of the 17th International Conference on Neural Information Pro...

  5. [13]

    Glazier, Jean-Paul Rieu, Arpita Upadhyaya, and Yasuji Sawada

    28 PREPRINT- JULY7, 2026 Noriyuki Bob Ouchi, James A. Glazier, Jean-Paul Rieu, Arpita Upadhyaya, and Yasuji Sawada. Improving the realism of the cellular potts model in simulations of biological cells.Physica A: Statistical Mechanics and its Applications, 329(3):451–458,

  6. [14]

    Kathryn M

    doi: 10.1109/TSMC.1979.4310076. Kathryn M. Ashman, Christine M. Bird, and Stephen E. Zepf. Detecting bimodality in astronomical datasets.The Astronomical Journal, 108:2348–2361,

  7. [15]

    Yungtai Lo

    doi: 10.1086/117248. Yungtai Lo. A likelihood ratio test of a homoscedastic normal mixture against a heteroscedastic normal mixture. Statistics and Computing, 18(2):233–240,

  8. [16]

    Joaquin Vanschoren, Jan N

    doi: 10.1007/s11222-008-9052-4. Joaquin Vanschoren, Jan N. van Rijn, Bernd Bischl, and Luis Torgo. OpenML: Networked science in machine learning. SIGKDD Explorations Newsletter, 15(2):49–60,

  9. [17]

    Bernd Bischl, Giuseppe Casalicchio, Matthias Feurer, Pieter Gijsbers, Frank Hutter, Michel Lang, Rafael G

    doi: 10.1145/2641190.2641198. Bernd Bischl, Giuseppe Casalicchio, Matthias Feurer, Pieter Gijsbers, Frank Hutter, Michel Lang, Rafael G. Mantovani, Jan N. van Rijn, and Joaquin Vanschoren. OpenML benchmarking suites. InAdvances in Neural Information Processing Systems (NeurIPS...

  10. [18]

    Gupta, Xiaojiang Chen, and Xin Wang Wang

    29 PREPRINT- JULY7, 2026 Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B. Gupta, Xiaojiang Chen, and Xin Wang Wang. A survey of deep active learning.ACM Computing Surveys (CSUR), 54(9):1–40,

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.