acceptodds
Under review as a conference paper at ICLR 2027

Folk Theorem-based Loss Function for Robust and Fair Learning under Class Imbalance

Abstract

Class imbalance biases deep classifiers toward majority classes and away from under-represented ones. We propose a loss function inspired by the Folk theorem from repeated-game theory: each class is treated, loosely, as a player whose classification accuracy is its payoff, and classes that persistently over-perform are assigned a dynamic penalty at both the batch level and the epoch level, while under-performing classes retain full gradient influence. The penalty is folded directly into each class's effective loss weight rather than added to the scalar loss value, so that it provably reshapes gradients instead of merely shifting a number (sec:method). We further combine per-class gradients into a Pareto-Optimal Gradient (POG) and reconcile that direction with the overall classification gradient via a two-objective PCGrad projection. The method is loss-agnostic and is evaluated on top of SCCE, Focal, LDAM, and Class-Balanced losses, across seven CNN backbones and four benchmark datasets (CIFAR-10, CIFAR-100, Fashion-MNIST, Caltech-101) under a single, unified preprocessing and imbalance protocol. We report macro-averaged precision/recall/F1, worst-class accuracy, and class-wise F1 variance alongside accuracy, and ablate the batch-level penalty, the epoch-level penalty, and the POG+PCGrad gradient fusion independently. Evaluated across seven CNN backbones and four base losses, the proposed penalty improves accuracy and fairness metrics over the corresponding baseline loss on the large majority of backbone/dataset combinations, most consistently when combined with SCCE. All code, configuration files, and raw per-run logs are released at an anonymized repository link: https://anonymous.4open.science/r/Folk-Theorem-Loss-3D8E

Primary area:
learning theory
Keywords:
Loss Function, Folk Theorem, Class Imbalance, Game Theory, Deep Learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_TP15OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a Folk-theorem inspired loss for fair learning, penalising classes with higher accuracy to even the playing field.

Critical strengths
  • The paper is upfront by framing the Folk-Theorem inspired contribution as a non formal result.
  • The game theoretic loss formulation is interesting in its own right.
Critical weaknesses
  • There are critical gaps in the literature. Results are used without attribution and related works with significant overlap in contributions are not acknowledged.
  • Reported results are inconsistent. Some entries in the results table are repeated. Results across tables are inconsistently reported or mixed up.
  • The reported results don't support the claims in the test. In Table 2 the proposed method underperforms baseline.
Questions for authors

Why are some equations used without attribution? Why are methods named in the text but not referenced? Can you explain the inconsistencies of the results in the tables?

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

Inspired by the Folk theorem in repeated games, this paper proposes a dynamic re-weighting loss that incorporates a batch-level penalty and an epoch-level penalty into each class's loss weight according to the class-wise accuracy observed during training. In addition, it introduces a gradient fusion step that uses PCGrad to resolve conflicts between the overall loss gradient and the Pareto-optimal gradient (POG), which is obtained by combining class-wise gradients in a min-norm manner. The authors claim that, across combinations of four datasets, seven CNN backbones, and four base losses (SCCE, Focal, LDAM, and CB), the proposed method generally improves accuracy and macro metrics.

Critical strengths
  • The motivation is well-founded, and the method takes the form of a simple plug-in.

  • The paper is candid about its limitations. It states from the outset that game theory serves only as an analogy and that no formal claims, such as those concerning Nash equilibria, are made. It also acknowledges that, since all results are from a single seed, differences of 0.5–1% are not significant. Furthermore, the authors themselves note that the comparison table with prior work does not constitute a controlled comparison, and they point out the computational cost of POG+PCGrad.

Critical weaknesses

W1. The proposed mechanism is inconsistent with the stated motivation, and its magnitude is too small to explain the reported performance gains. The main text describes the method as reducing the influence of over-performing classes, but Eq. 6 adds the non-negative terms PcP_c and FcF_c to the weight, so FcF_c in fact amplifies the gradients of classes with higher accuracy. In terms of magnitude, wcw_c ranges from approximately 0.55 to 5.5 at α=0.1\alpha=0.1, whereas Fc≤0.01F_c \le 0.01 and PcP_c changes by only ±0.001\pm 0.001 per batch. Moreover, with a batch size of 64, CIFAR-100 has fewer than one sample per class in each batch on average, so PcP_c is effectively a noise-driven random walk. It is therefore hard to accept that such marginal changes in weight alone raise the accuracy of MobileNetV2 on CIFAR-100 from 44.58 to 60.82, an improvement of nearly 16 points. In addition, POG+PCGrad improves F1 by only 0.01 points over Folk-only, while increasing the per-epoch training time on CIFAR-100 by roughly 19 times, so the practical benefit of the second contribution is unclear.

W2. The reliability of the results as a whole is questionable.

Since the imbalance is applied only to the training split, the test sets of CIFAR and Fashion-MNIST are balanced, in which case overall accuracy and macro recall should be identical. However, the two differ substantially in several cases, such as CIFAR-100 SCCE with Xception (Acc 64.31, Rec 48.91). In addition, macro F1 cannot exceed (macro P+macro R)/2(\text{macro P}+\text{macro R})/2, yet there are multiple violations of this bound; for instance, Focal+Folk with DenseNet201 on Caltech-101 reports P 87.93 and R 87.95 but F1 97.89. The numbers for the same setting also differ across tables. In particular, the MobileNetV2 CIFAR-100 SCCE baseline accuracy is 44.58 in Tables 1, 3, and 6 but 61.29 in Table 8, which means that, based on Table 8, the improvement of the proposed method is only 0.4 points. All results are from a single seed, and in the only fairness table (Table 2), the class-wise F1 standard deviation actually worsens from 5.79 to 6.16, which fails to support the "fair learning" claim in the title.

Furthermore, the positioning relative to prior work and the baseline comparisons are insufficient, making it difficult to assess the significance of the contribution. The paper claims that no existing method adapts to class-wise performance during training, but it does not discuss prior work that should be directly compared, such as CDB loss [R1], which updates weights based on class-wise accuracy, and Group DRO [R2], which increases weights according to worst-group performance. MGDA [R3] and PCGrad [R4], the original sources of POG and PCGrad, are also missing from the references. The compared methods are limited to LDAM and CB, while standard long-tail baselines such as LDAM-DRW, logit adjustment, Balanced Softmax, and decoupling (cRT) are omitted. Moreover, the paper relies solely on a non-standard protocol (75×75 resolution, a 7:1:2 split, and step imbalance), so its results cannot be compared with those reported on CIFAR-LT or ImageNet-LT. A fair comparison under the standard protocols commonly used by the above methods is necessary. The literature comparison in Table 10 lists methods unrelated to imbalance, such as Shadow loss, quantization, and edge enhancement, evaluated under different protocols, so it can hardly serve as supporting evidence. Finally, the provided link contains no code, so reproducibility is not ensured.

[R1] Sinha et al., Class-Wise Difficulty-Balanced Loss for Solving Class-Imbalance, ACCV 2020

[R2] Sagawa et al., Distributionally Robust Neural Networks, ICLR 2020

[R3] Sener & Koltun, Multi-Task Learning as Multi-Objective Optimization, NeurIPS 2018

[R4] Yu et al., Gradient Surgery for Multi-Task Learning, NeurIPS 2020

Questions for authors

Please refer to Critical Weaknesses

Reviewer_Wgt4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies class-imbalanced image classification and proposes a dynamically weighted training objective motivated by an analogy to the Folk theorem. The effective class weight combines a static inverse-frequency term with a batch-level state responding to changes in per-class accuracy and an epoch-level state reflecting differences in class accuracy. The authors also aggregate per-class gradients through a Pareto-style direction (POG) and fuse it with the loss gradient using PCGrad. Experiments span CIFAR-10, CIFAR-100, Fashion-MNIST and Caltech-101, seven CNN backbones and four underlying losses. Although the task is important and the evaluation is broad in scope, the central weighting equation and several reported results do not support the proposed mechanism as currently presented.

Critical strengths

S1. Relevant evaluation target and metrics. Class imbalance and minority-class reliability are important problems. The authors report macro-averaged metrics, worst-class accuracy and class-wise dispersion alongside aggregate accuracy, which is more informative than overall accuracy alone.

S2. Breadth and component-level reporting. The study tests multiple CNN backbones and base losses and includes ablations and computational-overhead measurements. It also appropriately treats the Folk theorem as an analogy rather than claiming to establish a game-theoretic equilibrium theorem.

Critical weaknesses

W1. The proposed penalty mechanism conflicts with its equations (Eqs. 4–6). Equation (6) specifies w_eff,c = max(w_c + P_c + F_c, 0). With positive static weights and nonnegative penalty states, this implies w_eff,c ≥ w_c. Therefore, the displayed rule cannot reduce an over-performing class’s weight below its static baseline as repeatedly claimed. Moreover, the epoch term F_c rises with class accuracy, while the batch update increases P_c after a decrease in accuracy, not after persistent over-performance. There is also an internal normalization inconsistency: substituting ACC_c = b into Eq. (5) gives F_c(b) = (b − a)λ_f, whereas the text and Appendix A.1 describe an endpoint of λ_f. The precise rule implemented in the experiments is consequently unclear. These are central methodological concerns, not merely presentation defects.

W2. Inconsistent experimental reporting and insufficient evidence for the added gradient-fusion cost. For CIFAR-100/MobileNetV2/SCCE, the reported baseline accuracy differs between Tables 1/3 (44.58%) and Table 8 (61.29%); corresponding full-method accuracy is 60.83% versus 61.69%. The Fashion-MNIST/Xception result also differs between Tables 6 and 10 (97.74% versus 91.39%). Separately, Table 2 shows class-wise F1 standard deviation increasing from 5.79 to 6.16 despite the broad fairness claim. In Table 3, POG+PCGrad improves accuracy over Folk-only by only 0.01 percentage points (60.82% to 60.83%), while Table 4 indicates substantial additional training cost. Without reconciled tables, multiple seeds and a compute-matched ablation, neither the effect size nor the benefit of the costly fusion step is convincing.

Questions for authors

Q1. Is the plus sign in Eq. (6) exactly what was implemented? If so, how can adding the nonnegative P_c and F_c terms suppress the effective contribution of consistently strong classes? Please show class-weight and gradient-norm trajectories.

Q2. Which formula for F_c was used, Eq. (5) or the endpoint-normalized definition in Appendix A.1? How are unchanged class accuracy and classes absent from a mini-batch handled?

Q3. Can the authors provide one auditable, run-level reconciliation of the conflicting values in Tables 1, 3, 6, 8 and 10, with metric definitions and random seeds?

Q4. Across at least three seeds, does POG+PCGrad improve worst-class accuracy or macro F1 over the penalty-only method under matched wall-clock training cost?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.