A Coreset Selection Framework with Ensemble Aggregation for Image Classification

Lucas Pascotti Valem; Pedro Rocha Dantas

REVIEW 2 major objections 5 minor 14 references

A Coreset Selection Framework with Ensemble Aggregation for Image Classification

T0 review · 2 major / 5 minor · reviewed 2026-07-13 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Stratified sampling over score intervals plus ensemble aggregation yields competitive coresets that cut SGC training cost while preserving accuracy.

desk verdict Solid undergraduate workshop paper: SCOSS_B is a clean stratified recipe that often ranks best for SGC on two datasets, with real SGC cost savings and public code; scope and fixed centroid score keep the claim modest, not broken. read the letter →

arxiv 2607.09100 v1 pith:RQQMS6AD submitted 2026-07-10 cs.CV cs.AI

Pedro Rocha Dantas , Lucas Pascotti Valem This is my paper

classification cs.CVcs.AI

keywords coresetselectionstratifiedsamplingensembleaggregationimageclassificationSimpleGraphConvolutionclass-balanceddataefficiency

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

Large image datasets make full-set training expensive in time and memory, especially for graph-based classifiers that use the whole graph. The paper argues that a good coreset can be built by first scoring every training image (here, Euclidean distance of a deep feature to its class centroid), then drawing samples from equal-sized intervals that cover the full score range rather than only the middle or random points. Class-balanced and graph-aware variants of this idea are introduced, and predictions from several independent coresets are averaged to reduce run-to-run noise. On a fine-grained bird dataset and on CIFAR-10, the class-balanced version is competitive with or better than random and moderate baselines for Simple Graph Convolution, often ranking first, and delivers large cuts in training time and GPU memory at modest accuracy cost. The practical claim is that covering the whole score distribution with stratified, class-aware sampling plus simple ensembling is enough to obtain usable accuracy-efficiency trade-offs without more elaborate coreset machinery.

What carries the argument

SCOSS: sort training samples by a score (distance to class centroid), partition into equal-count intervals that cover the full score range, and draw a proportional random sample from every interval; class-balanced and connected-component variants allocate the budget per class or exploit a kNN graph.

What would settle it

Replace the centroid-distance score with an uninformative or adversarial score (or recompute scores after each selection) and check whether SCOSS still ranks above random and Moderate Coreset for SGC accuracy at the same sampling ratios on CUB-200 and CIFAR-10.

Watch

Extended reading notes

Core claim

SCOre-Stratified Selection (SCOSS), especially its class-balanced form, produces training subsets that are competitive with random and Moderate Coreset baselines and frequently achieve the best average rank for Simple Graph Convolution across sampling ratios on CUB-200 and CIFAR-10, while substantially reducing SGC training time and GPU memory; on the fine-grained set the same coresets let SGC outperform SVM when few labels are used.

Load-bearing premise

The Euclidean distance of a fixed ResNet feature vector to its class centroid is assumed to be a good enough, once-and-for-all ranking of sample usefulness for later stratified selection.

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.

Desk Editor's Note

read the letter

This is a clean, usable empirical methods paper from an undergrad project, not a theory or foundation-model result. The one thing worth knowing is that class-balanced SCOSS_B frequently takes the best average rank for SGC across ratios on CUB-200 and CIFAR-10 (Tables 3–4), while Table 2 documents large SGC time/memory cuts at 20% sampling (~74% time, ~64% memory on CIFAR-10) with usable accuracy trade-offs. Ensembles help when run variance is high.

What is actually new is the concrete recipe: equal-count score bins over the full distribution, class-balanced budgets with Mc proportional to score range, plus a connected-component variant, wrapped in multi-run probability averaging. The pieces (stratified/coverage sampling, class-proportional allocation, Moderate Coreset) are cited; the assembly and the SGC-focused comparison are the contribution. They do the basics well: mean±std over 10 runs and 5 ensembles, full-set baselines, multiple ratios, two classifiers, average ranks, and public code. Claims stay close to the tables. No circularity—score is fixed external centroid distance, evaluation is held-out accuracy.

Soft spots are real but proportional. The load-bearing choice is Euclidean distance to class centroid computed once on the full training set and never updated; if that score is a poor proxy for utility, stratified coverage of it still only beats random and median-only Moderate under that score. No score ablation, no other modern coreset baselines, fixed M=16/p=0.9/ResNet-152, only two datasets and two classical models. On CIFAR-10 differences shrink and SVM often leads; the “SGC with SCOSS can beat SVM at low labels” point is mainly the fine-grained set. None of that invents a contradiction with the reported numbers—the ranking claim is comparative and modest, and the stress-test is right that it holds under the stated score.

Who it is for: people who train SGC/transductive models or need simple, reproducible coreset heuristics with cost numbers. Not for anyone looking for a new theory of sample importance or large-scale foundation results. Math is elementary; data and citations are honest for a workshop paper. I would send it to peer review at a workshop or short-paper track; it deserves a referee, not a desk reject. Engage if you care about cheap SGC coresets; skip if you need broader validation first.

Referee Report

2 major / 5 minor

Summary. The paper proposes a coreset-selection framework for image classification that pairs a new stratified sampler (SCOSS) with optional multi-run ensemble aggregation. SCOSS sorts training samples by a fixed score (Euclidean distance of ResNet-152 features to the class centroid), partitions them into equal-count intervals, and draws a budgeted sample from each interval; class-balanced (SCOSS_B) and graph-component (SCOSS_CC) variants are also defined. Baselines are random and Moderate Coreset, each with and without class balancing. The selected coresets train SGC and SVM at sampling ratios 2.5–20 % on CUB-200 and CIFAR-10; accuracy is reported with and without probability averaging over 10 independent runs. Tables 1–4 show that SCOSS_B frequently obtains the best average rank for SGC, that ensemble aggregation reduces run-to-run variance, and that 20 % sampling yields large SGC time/memory savings (e.g., ~74 % time and ~64 % memory on CIFAR-10) at a usable accuracy cost. Code and supplementary materials are released.

Significance. If the reported rankings hold under broader conditions, the work supplies a simple, reproducible stratified coreset recipe that is competitive with standard random and median-focused baselines for graph-based image classifiers, together with concrete efficiency numbers that matter for resource-constrained or transductive settings. Strengths that raise the contribution above a pure engineering note include: public code, multi-run mean±std reporting, explicit full-set baselines, and a transparent average-rank summary that supports the modest “often best for SGC” claim. The contribution remains incremental—limited to two datasets, two classifiers, and a single fixed score—but is useful for practitioners who already rely on SGC-style models and need a lightweight selection heuristic.

major comments (2)

§2.1–2.2 and Tables 3–4: the central comparative claim rests on a single, never-updated score (Euclidean distance to class centroid). No ablation of alternative scores (e.g., model uncertainty, gradient norms, or graph-based centrality) is provided. Because the paper’s own conclusions section lists “testing alternative scoring criteria” as future work, the present ranking of SCOSS_B versus Moderate/Random is only guaranteed under this particular score; a short score-sensitivity experiment (or an explicit limitation statement that the ranking is score-conditioned) is needed before the claim can be treated as general.
§3 and Tables 3–4: only random and Moderate Coreset (plus class-balanced variants) are used as baselines. More recent stratified or coverage-centric methods cited in the introduction (Zheng et al. 2023; Tsai et al. 2025) are not re-implemented. Without at least one stronger contemporary baseline, the statement that SCOSS is “competitive with existing methods” is under-supported; adding one such baseline (or clearly restricting the claim to the two classical baselines) is load-bearing for the paper’s positioning.

minor comments (5)

§2.2: the formulas for Mc and the remainder-filling step are clear, but the interaction between rounding of k and the final uniform top-up is never illustrated with a small numerical example; a one-line example would remove ambiguity.
Table 2: the 2.5 % row for CUB-200 is left blank (“–”); a brief note explaining why that ratio was omitted (memory? class-balance failure?) would help readers.
Figure 2 caption and surrounding text: the density sketch is helpful, yet the figure itself is not referenced when the equal-count interval construction is first introduced; a forward pointer would improve flow.
§3: hyper-parameters M=16 and p=0.9 are fixed without sensitivity analysis; even a short remark that results were stable for M∈{8,16,32} would strengthen reproducibility claims.
References: several arXiv preprints (Moser et al. 2026, Tsai et al. 2025) lack venue or version information; standardizing them would aid archival stability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical coreset comparison with fixed external score, independent baselines, and held-out test accuracy; nothing reduces by construction to its inputs.

full rationale

This is a standard empirical methods paper. SCOSS is defined as stratified sampling over a fixed, once-computed score (Euclidean distance of ResNet-152 features to class centroid, §2.1–2.2); coresets are then used to train SGC/SVM and evaluated by accuracy on the full held-out test set (Tables 3–4). Baselines (Random, Moderate Coreset, and class-balanced variants) are independent algorithms from the literature or uniform sampling; they are not derived from SCOSS. The ensemble is simple probability averaging over independent runs and does not force rankings. There is no fitted parameter that is later called a prediction, no uniqueness theorem imported from the authors, no self-citation that carries the central claim, and no renaming of a known result as a derivation. The modest comparative claim (SCOSS_B often best average rank for SGC under this score) is supported by the reported numbers and is not true by construction. Score 0 is the correct outcome.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central accuracy/efficiency claims rest on standard ML practice plus a handful of hand-set design knobs and the unproved modeling choice that centroid distance is a useful importance score. No new physical entities; the invented objects are algorithmic constructs defined by the paper.

free parameters (4)

Number of score intervals M
Fixed to M=16 for all SCOSS variants with no sensitivity study; controls how finely the score axis is stratified and directly affects which samples can be drawn.
SCOSS_CC budget split p
p=0.9 allocates most of the coreset to stratified class-balanced sampling and the rest to kNN components; chosen by hand.
kNN graph parameters for SGC and SCOSS_CC
Reciprocal kNN k=40/30 for SGC and k=5 for SCOSS_CC, plus learning rates and epoch counts, are dataset-specific fixed settings that influence both selection and classifier performance.
Score function (Euclidean distance to class centroid)
Adopted as the sole scoring criterion without comparison to alternatives the conclusions themselves list as future work; the entire stratification is defined on this scalar.

assumptions (4)

domain assumption ResNet-152 ImageNet features are a fixed, adequate representation for coreset scoring and for SGC/SVM classification on CUB-200 and CIFAR-10.
Feature extraction is performed once with a pretrained network (§2.1, §3); all subsequent selection and training operate only in that embedding.
domain assumption Samples near the class centroid in feature space are more 'representative,' and covering the full score distribution improves coreset quality versus median-only selection.
Motivates both the score definition and SCOSS versus Moderate Coreset (§1–2.2); not derived, only empirically tested.
domain assumption Simple probability averaging over independent coreset runs is a valid ensemble aggregation for reducing selection variance.
Stated in §2.1 with a general ensemble citation; form of aggregation is not ablated.
standard math Standard definitions of random sampling, Moderate Coreset, SGC, and RBF-SVM as implemented.
Baselines and classifiers are taken from cited literature and used as black-box comparators.

invented entities (1)

SCOSS / SCOSS_B / SCOSS_CC selection procedures
purpose: Define how training indices are chosen from score intervals, per-class budgets, and optional kNN connected components.
Algorithmic constructs introduced in §2.2; evaluated only inside this paper's experimental loop. independent_evidence is false because no external replication or theoretical guarantee is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Coreset Selection Framework with Ensemble Aggregation for Image Classification." pith.science (2026). https://pith.science/paper/RQQMS6AD

@misc{pith2026260709100,
  author       = {Pith},
  title        = {Pith review of: A Coreset Selection Framework with Ensemble Aggregation for Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RQQMS6AD}},
  note         = {Machine review of arXiv:2607.09100}
}

read the original abstract

The rapid growth of image data has produced large-scale datasets, raising concerns about the time and memory costs of model training. Selecting representative training subsets, however, remains challenging: individual sample contributions are unclear, and model behavior varies across datasets and runs. We address these challenges with a framework that combines coreset selection with an ensemble aggregation over multiple runs. For coreset selection, we propose SCOre-Stratified Selection (SCOSS), which partitions the training data into intervals based on a chosen score and samples from each interval. The ensemble combines predictions from multiple runs, each performed on an independently sampled training subset. As baselines, we use moderate and random selection, each in original and class-balanced versions. We assess the framework with Simple Graph Convolution (SGC) and Support Vector Machine (SVM) classifiers under different sampling ratios. Experiments show that SCOSS is competitive with baselines, often the best choice for SGC, and enables favorable trade-offs between accuracy and efficiency. On the fine-grained dataset, SGC with SCOSS outperforms SVMs when using fewer labeled samples. The code and supplementary materials are publicly available at http://scoss.lucasvalem.com.

Figures

Figures reproduced from arXiv: 2607.09100 by the authors.

**Figure 2.** Visual Representation of SCOSS interval division. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 1 canonical work pages

[1]

Survey on Graph Neural Networks , year=

Gkarmpounis, Georgios and Vranis, Christos and Vretos, Nicholas and Daras, Petros , journal=. Survey on Graph Neural Networks , year=
[2]

arXiv preprint arXiv:2505.17799 , year=

A Coreset Selection of Coreset Selection Literature: Introduction and Recent Advances , author=. arXiv preprint arXiv:2505.17799 , year=. 2505.17799 , archivePrefix=

arXiv
[3]

2022 , isbn =

Guo, Chengcheng and Zhao, Bo and Bai, Yanbing , title =. 2022 , isbn =. doi:10.1007/978-3-031-12423-5_14 , booktitle =

work page doi:10.1007/978-3-031-12423-5_14 2022
[4]

arXiv preprint arXiv:1609.02907 , year=

Semi-supervised classification with graph convolutional networks , author=. arXiv preprint arXiv:1609.02907 , year=

arXiv
[5]

Core‐sets: An updated survey , volume =

Feldman, Dan , year =. Core‐sets: An updated survey , volume =. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery , doi =
[6]

arXiv preprint arXiv:2507.10904 , year=

Class-Proportional Coreset Selection for Difficulty-Separable Data , author=. arXiv preprint arXiv:2507.10904 , year=. 2507.10904 , archivePrefix=

arXiv
[7]

arXiv preprint arXiv:2210.15809 , year=

Coverage-centric Coreset Selection for High Pruning Rates , author=. arXiv preprint arXiv:2210.15809 , year=. 2210.15809 , archivePrefix=

arXiv
[8]

ICLR , year=

Moderate Coreset: A Universal Method of Data Selection for Real-world Data-efficient Deep Learning , author=. ICLR , year=

Show all 14 references

[9]

Simplifying Graph Convolutional Networks

Wu, Felix and Souza, Amauri and Zhang, Tianyi and Fifty, Christopher and Yu, Tao and Weinberger, Kilian. Simplifying Graph Convolutional Networks. Proceedings of the 36th International Conference on Machine Learning. 2019

2019
[10]

Machine Learning , volume =

Corinna Cortes and Vladimir Vapnik , title =. Machine Learning , volume =. 1995 , publisher =

1995
[11]

WIREs Data Mining and Knowledge Discovery , volume =

Sagi, Omer and Rokach, Lior , title =. WIREs Data Mining and Knowledge Discovery , volume =. doi:https://doi.org/10.1002/widm.1249 , url =. https://wires.onlinelibrary.wiley.com/doi/pdf/10.1002/widm.1249 , abstract =

doi
[12]

and Zhang , X

He , K. and Zhang , X. and Ren , S. and Sun , J. Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2016

2016
[13]

Learning multiple layers of features from tiny images , year =

Krizhevsky, Alex and Hinton, Geoffrey , address =. Learning multiple layers of features from tiny images , year =
[14]

and Branson, S

Wah, C. and Branson, S. and Welinder, P. and Perona, P. and Belongie, S. , title =. 2011 , institution =

2011

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.

[1] [1]

Survey on Graph Neural Networks , year=

Gkarmpounis, Georgios and Vranis, Christos and Vretos, Nicholas and Daras, Petros , journal=. Survey on Graph Neural Networks , year=

[2] [2]

arXiv preprint arXiv:2505.17799 , year=

A Coreset Selection of Coreset Selection Literature: Introduction and Recent Advances , author=. arXiv preprint arXiv:2505.17799 , year=. 2505.17799 , archivePrefix=

arXiv

[3] [3]

2022 , isbn =

Guo, Chengcheng and Zhao, Bo and Bai, Yanbing , title =. 2022 , isbn =. doi:10.1007/978-3-031-12423-5_14 , booktitle =

work page doi:10.1007/978-3-031-12423-5_14 2022

[4] [4]

arXiv preprint arXiv:1609.02907 , year=

Semi-supervised classification with graph convolutional networks , author=. arXiv preprint arXiv:1609.02907 , year=

arXiv

[5] [5]

Core‐sets: An updated survey , volume =

Feldman, Dan , year =. Core‐sets: An updated survey , volume =. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery , doi =

[6] [6]

arXiv preprint arXiv:2507.10904 , year=

Class-Proportional Coreset Selection for Difficulty-Separable Data , author=. arXiv preprint arXiv:2507.10904 , year=. 2507.10904 , archivePrefix=

arXiv

[7] [7]

arXiv preprint arXiv:2210.15809 , year=

Coverage-centric Coreset Selection for High Pruning Rates , author=. arXiv preprint arXiv:2210.15809 , year=. 2210.15809 , archivePrefix=

arXiv

[8] [8]

ICLR , year=

Moderate Coreset: A Universal Method of Data Selection for Real-world Data-efficient Deep Learning , author=. ICLR , year=