acceptodds
Under review as a conference paper at ICLR 2027

GCorP: Global Correlation Pooling

32 views

Abstract

Global covariance pooling summarizes how channel responses vary together. However, it also encodes each channel's scale. Rescaling a channel can therefore change the representation even when the channel relationships are unchanged. We introduce Global Correlation Pooling (GCoRP), which normalizes covariance into a scale-invariant correlation matrix and maps this matrix to geometry-specific flat or hyperbolic coordinates. For flat geometries, we show that non-trivialized correlation-manifold multinomial logistic regression can be optimized exactly in Euclidean coordinates: Riemannian SGD on manifold-valued prototypes and tangent vectors maps to ordinary Euclidean SGD under the induced coordinate map, without tangent-space trivialization. We evaluate GCoRP on image classification, EEG foundation-model classification, and generative-model post-training. Across these tasks, GCoRP offers a favorable accuracy–efficiency trade-off and supports both classification and distribution-matching training.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
representative learning, multimodal, image-generation, post-training

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ewQ4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Global Correlation Pooling (GCorP), which constructs correlation matrices from backbone features and maps them through flat or hyperbolic correlation geometries. It also derives a Euclidean realization of flat CorMLR under specific parameterization and optimization conditions. Experiments on image classification, EEG classification, and generative-model post-training report competitive performance with lower computational cost in several settings.

Critical strengths
  • The fixed-MPA covariance-versus-correlation comparison provides direct evidence that variance standardization can improve recognition under otherwise matched settings.

  • The evaluation covers three distinct applications. The EEG experiments use shared initialization and validation-based configuration selection, with results reported across repeated runs.

Critical weaknesses
  • The Poincaré and Lorentz variants use β-Concat and H-Concat, which can yield different distances for the same pair of correlation matrices. The Lorentz–Poincaré isometry between individual factors does not establish equivalence after these different aggregation operations, undermining the interpretation of the two post-training objectives as realizations of the same distance.

  • Its optimization-equivalence result concerns SGD on paired variables (q, a), whereas ImageNet training uses AdamW. OLM and LSM also use approximate logarithms, and LSM uses one Newton step for diagonal scaling. Although the numerical sweeps support predictive performance, the geometric fidelity of these implemented heads remains unquantified, leaving the connection between the exact theory and empirical results incomplete.

  • All GCorP losses retain the mean term, and no mean-only result is reported. The comparison with exact covariance FD supports the combined objective, but does not establish the correlation term’s incremental contribution.

Questions for authors

see weakness

Reviewer_6s69OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GCorP, which converts global covariance representations into correlation matrices and applies geometry-specific flat or hyperbolic mappings. It further establishes a Euclidean realization of flat correlation-manifold logistic regression and evaluates the framework across image classification, EEG decoding, and generative-model post-training.

Critical strengths
  1. The paper provides a unified framework for correlation-based second-order representations across multiple learning tasks.
  2. The theoretical analysis offers an explicit connection between flat correlation-manifold classification and Euclidean optimization.
  3. Experiments cover multiple backbones and applications, with GCorP-OLM achieving competitive accuracy and favorable computational efficiency.
Critical weaknesses
  1. The necessity of the proposed geometric transformations is not sufficiently established. Although Table 8 demonstrates that replacing covariance with correlation improves the MPA baseline, this does not establish the additional benefit of geometry-specific mappings. A direct correlation-vectorization baseline followed by an FC or MLP classifier is missing. Since CoRNet (ICML 2026) already introduces the underlying correlation-manifold geometries, the paper needs to isolate what the proposed pooling construction contributes beyond correlation normalization and existing geometric transformations.
  2. The theoretical equivalence has limited demonstrated implications for optimization. Proposition 4.1 establishes equivalence by equipping the parameter space with a pullback metric specifically induced by the chosen Euclidean coordinates. While mathematically meaningful, this equivalence largely follows from the coordinate construction rather than establishing an independent optimization advantage. Moreover, the equivalence concerns Euclidean optimization over the induced parameter pairs, not unconstrained SGD on ordinary FC weights and biases. The experiments do not demonstrate whether this theoretical formulation provides advantages over simpler Euclidean classifiers.
  3. The claimed equivalence between the two hyperbolic constructions appears mathematically inconsistent. Appendix E.4 states that the Poincaré and Lorentz PHCM distances agree because their representations are related by an isometry. However, the Poincaré construction uses dimension-dependent β\beta-concatenation, whereas the Lorentz construction directly concatenates spatial components. These operations do not generally define the same embedding. For example, with a 3×33\times3 correlation matrix having only C12=C21=0.5C_{12} = C_{21} = 0.5 as nonzero off-diagonal entries, β\beta-concatenation scales the nontrivial row's distance by β3/β1=1/2\beta_3/\beta_1=1/2, while the Lorentz construction preserves its original distance from the identity. Thus, the asserted distance equivalence requires correction or additional justification.
  4. The empirical results do not demonstrate consistent benefits across the proposed geometries. Table 2 shows that OLM is competitive, but several other GCorP variants perform substantially worse, particularly ECM and LECM on ViT-B/16. On ResNet-101, MPA still slightly outperforms GCorP-OLM in Top-1 accuracy (80.38% versus 80.31%). Similarly, the generative-model results largely match rather than improve upon exact covariance-based training. The paper does not sufficiently explain why certain geometries succeed while others fail, making the claimed general advantages of correlation-manifold pooling less convincing.
Questions for authors

See weakness.

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

This paper introduce a novel second-order representation method, i.e., Global Correlation Pooling (GCorP). Several instantiations, e.g., ECM, LECM, OLM, LSM, and PHCM, are further provided. Experiments on image recognition, EEG classification, and generative-model posttraining are conducted to evaluate the proposed methods. However, this work suffers from several limitations, e.g., (i) less improvement against lobal covariance pooling, (ii) insufficient evaluation, and the reviewer does not recommend acceptance.

Critical strengths
  • A novel second-order representation method, i.e., Global Correlation Pooling (GCorP).
  • Several instantiations, e.g., ECM, LECM, OLM, LSM, and PHCM.
  • Exact Euclidean realization of non-trivialized correlation-manifold multinomial logistic regression.
Critical weaknesses
  • More explanations on the motivation is required. What motivates the authors to use Correlation Pooling instead of covariance pooling? It is claimed that Correlation Pooling is an alternative of covariance pooling, which is not sufficient.
  • From Tables 2&3, the performance improvements against the counterparts are marginal.
  • The generative model post-training in Sec. 6.2 actually is a distillation experiment instead of reinforcement learning.
  • Moreover, for all the experiments, instead of comparing with the baselines, it is also suggested to compare with SoTAs of these tasks.
Questions for authors

Please refer to the weakness.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.