acceptodds
Under review as a conference paper at ICLR 2027

GEOMETRY-PRESERVING STRUCTURED PRUNING FOR LONG-TAILED RECOGNITION

Abstract

Long-tailed recognition models are trained on highly imbalanced data, where tail classes receive limited sample support. Structured channel pruning based on cross-sample importance may discard channels supporting tail-class regions, distorting local geometry and inter-class separation and thereby re-amplifying class bias during compression. To this end, we formulate long-tailed pruning as the preservation of class-local feature-space geometry and propose GB-DCS (Granular-Ball D-Optimal Channel Subspace Selection). GB-DCS uses class-specific granular balls to characterize local representation regions and temporarily masks candidate channels to measure signed deformations in their position, compactness, and separation from other classes. After balancing these responses across classes and frequency groups, GB-DCS applies a multi-kernel D-optimal criterion to retain channel subspaces with complementary geometric responses under a given structural budget. Experiments on moderate-scale CIFAR-100-LT and large-scale ImageNet-LT show improved overall recognition compared with existing methods, with substantial gains on tail classes. GB-DCS provides a geometry-aware framework for balancing model compactness and class-local representation preservation under long-tailed imbalance.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Long-tailed recognition, Feature geometry preservation, Structured channel pruning, Granular-ball computing, Performance‑overhead trade‑off

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_2LRtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies structured channel pruning for recognition under class imbalance, with an emphasis on preserving representations of rare classes. The proposed method, GB-DCS, models local class regions using granular balls, measures geometric changes caused by masking individual channels, and combines class balancing with a determinant criterion to retain complementary channels. Experiments on CIFAR-100-LT and ImageNet-LT show improvements over the reported pruning baselines, particularly for tail classes in several settings.

Critical strengths
  1. The paper addresses the meaningful problem of disproportionate accuracy loss on rare classes during model compression.
  2. Local geometric responses provide an intuitive basis for preserving class representations, while complementary selection accounts for channel redundancy.
  3. Experiments span two benchmarks and three training objectives, with substantial tail accuracy gains in some settings at matching FLOPs budgets.
  4. The paired ablations and geometry diagnostics provide useful evidence for the contributions of the main components.
Critical weaknesses
  1. Selecting configurations and checkpoints using test performance may lead to optimistic results. Validation based selection would improve reliability.
  2. Differences in training budgets and classifier adaptation make it difficult to isolate the benefits of the pruning criterion.
  3. The connection between determinant maximization and geometry preservation under joint pruning remains heuristic, with limited analysis of channel interactions.
  4. Selection time and memory usage are not reported, leaving the practical cost of evaluating individual channel masks unclear.
Questions for authors

Please refer to weaknesses.

Reviewer_BMWKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GB-DCS to prune channels by tracking how individual masks change class-local feature geometry. It balances these responses across classes and uses a multi-kernel log-det objective to keep complementary channels. Results on CIFAR-100-LT and ImageNet-LT show promising tail-class gains, particularly with LDAM on ImageNet-LT.

Critical strengths
  1. This method addresses a real weakness of aggregate pruning scores: minority class regions can get overlooked. Combining local responses with channel complementarity is a sensible design.

  2. This paper provides a useful controlled ablation. The diagonal-kernel comparison, in particular, isolates the value of channel relationships rather than just individual response strength.

Critical weaknesses
  1. The main results are hard to interpret cleanly. For example, Section 4.1 denotes checkpoints/configurations are selected by test performance and budgets differ across methods. Classifier adaptation and BN settings also vary.

  2. Selecting complementary channels does not consistently improve overall accuracy over simply keeping channels with the largest response energies. More runs with different random seeds are needed to check whether this extra step reliably benefits overall and tail-class accuracy.

Questions for authors
  1. All experiments use ResNet-32/50. What changes would GB-DCS need for a non-ResNet backbone, particularly in the feature stages and pruning dependency groups?

  2. Beyond ImageNet-LT, have you evaluated naturally long-tailed, fine-grained data such as iNaturalist? I am particularly interested in whether granular-ball estimates remain reliable with very few tail-class examples, and whether tail accuracy improves at a comparable pruning budget.

  3. This paper provides useful single-factor sensitivity tests, but only for CIFAR-100-LT/ResNet-32 at IR=100, while Table 7 changes several settings across datasets. How were these hyper-parameters selected, and can one fixed configuration transfer across datasets or imbalance ratios without retuning?

Reviewer_kdCzOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies structured channel pruning for long-tailed recognition. It argues that importance-based pruning can discard channels that support tail classes, and frames the problem as preserving class-local feature geometry. The proposed GB-DCS models each class with local “granular balls,” measures how removing each channel distorts this geometry, balances the responses across classes, and selects complementary channels with a D-optimal criterion. Experiments on CIFAR-100-LT and ImageNet-LT with ResNet backbones report improvements, particularly on tail classes.

Critical strengths
  1. Treating long-tailed pruning as preserving class-local feature geometry, rather than ranking channels by aggregate importance, is a reasonable perspective. Signed responses and sample-disjoint search/guard folds are a sensible way to measure it.

  2. The method is described in enough detail that a reimplementation appears feasible, and the discussion section is candid about the single-candidate approximation and fixed quota limitations.

Critical weaknesses
  1. The contributions of this paper are limited. The method combines existing components: granular-ball modeling, class-wise reweighting, DPP/log-determinant subset selection, and DepGraph-based pruning. It does not offer a new insight into why or how pruning harms tail classes.

  2. The writing of this paper needs to be further improved. Specifically, the logic of the paper is hard to follow. The method section is notation-heavy and lacks a clear high-level narrative linking the problem formulation to the optimized objective. Important experimental details, such as how models and configurations were selected, are not presented where the main results appear.

  3. The paper does not adequately position itself against recent advances in long-tailed recognition from the past two years. The training methods and compared baselines are dated and few, which makes it hard to judge whether the approach is competitive with the current state of the art.

  4. The method is formulated and evaluated only on CNNs (ResNets). It is unclear how it would extend to Transformer-based architectures such as ViTs, which are increasingly standard in long-tailed recognition.

  5. The evaluation covers only CIFAR-100-LT and ImageNet-LT. Widely used large-scale, real-world benchmarks such as iNaturalist 2018 (and Places-LT) are missing. This limits conclusions about scalability and generality.

Questions for authors
  1. How does the method compare with recent long-tailed recognition methods and stronger pruning baselines?

  2. Can the method be extended to Transformer-based backbones, and how would it perform there?

  3. How does GB-DCS perform on iNaturalist 2018 or Places-LT?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.