Research map / Label-efficient and robust learning / Recognition and long-tail
SCoRe: Submodular Combinatorial Representation Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SCoRe, a family of submodular combinatorial loss functions (Total Information and Total Correlation with facility-location, graph-cut, and log-determinant instantiations) for long-tail representation learning that generalizes SupCon, N-pairs and OPL.
Problem
Long-tail visual recognition suffers from inter-class bias and intra-class variance; existing metric/contrastive losses such as SupCon do not adequately penalize inter-class bias.
Approach
Treat each class as a set and define losses from submodular information measures (Total Submodular Information S_f and Total Correlation C_f) over pairwise cosine-similarity kernels, yielding instantiations (SCoRe-FL, SCoRe-GC, SCoRe-LogDet) that jointly reduce intra-class variance and inter-class bias and are class-balancing in the FL case.
Key results
- SCoRe-FL improves PaCo on CIFAR-10-LT IF=100 to 85.61% and GLMC+SCoRe-FL to 92.33% (vs 88.50 GLMC), with up to 7.6% gain over SupCon on the step distribution.
- GLMC+SCoRe-FL reaches 59.3% All-accuracy on ImageNet-LT vs 56.1 for GLMC (+3.2 points), and PaCo+SCoRe-FL reaches 57.5 vs 55.6.
- On object detection, Faster-RCNN+SCoRe-FL achieves 36.3 mAP on IDD (vs 30.4 baseline) and 19.1 mAP on LVIS (vs 14.2 baseline).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SCoRe-FL to GLMC | GLMC · CIFAR-10-LT IF=100 | Top-1 accuracy: 88.50 -> 92.33 | Helped |
| Add SCoRe-FL to GLMC | GLMC · CIFAR-100-LT IF=100 | Top-1 accuracy: 58.0 -> 61.33 | Helped |
| Add SCoRe-FL to GLMC | GLMC · CIFAR-10-LT IF=10 | Top-1 accuracy: 94.90 -> 94.93 | Helped |
| Add SCoRe-FL to PaCo | PaCo · ImageNet-LT All | Top-1 accuracy: 55.6 -> 57.5 | Helped |
| Use SCoRe-FL instead of SupCon | ResNet-50 representation learning · CIFAR-10-LT Step | Top-1 accuracy: 78.10 -> 87.49 | Helped |
| Use SCoRe-FL instead of SupCon | ResNet-50 representation learning · DermaMNIST | Top-1 accuracy: 72.12 -> 73.77 | Helped |
| Use SCoRe-FL in the box classification head | Faster-RCNN + FPN · IDD | mAP: 31.2 -> 36.3 | Helped |
| Use SCoRe-FL in the box classification head | Faster-RCNN + FPN · LVIS v1.0 | mAP: 14.4 -> 19.1 | Helped |
| Set λ to 0.5 instead of 1.0 (ablation) | SCoRe-GC · CIFAR-10-LT LongTail | Top-1 accuracy: 89.96 -> 83.65 | Hurt |
| Set λ to 1.5 instead of 1.0 (ablation) | SCoRe-GC · CIFAR-10-LT LongTail | Top-1 accuracy: 89.96 -> 87.11 | Hurt |
| Set λ to 2.0 instead of 1.0 (ablation) | SCoRe-GC · CIFAR-10-LT LongTail | Top-1 accuracy: 89.96 -> 85.86 | Hurt |
Surprising findings
- SCoRe-GC with PaCo underperforms PaCo on ImageNet-LT Few and Med shots (16.7 Few vs 34.9 for PaCo), giving a lower overall All accuracy (50.2 vs 55.6).
- SCoRe-FL yields lower Few-shot accuracy on ImageNet-LT despite better overall accuracy, which the authors attribute to tail overfitting.
Limitations
- Gains are mixed across settings: SCoRe-GC underperforms PaCo on ImageNet-LT tail classes, and SCoRe-FL lowers Few-shot accuracy.
- Balanced-setting accuracy of SCoRe objectives is lower than SoTA contrastive learners in the stage-1 epoch ablation.
- SCoRe-GC submodularity requires lambda >= 1, and the lambda choice is tuned on CIFAR-10 LT.
Experimental setup
Datasets: CIFAR-10-LT CIFAR-100-LT ImageNet-LT OrganAMNIST DermaMNIST IDD LVIS v1.0 CIFAR-10
Benchmarks: CIFAR-10-LT CIFAR-100-LT ImageNet-LT MedMNIST IDD LVIS v1.0
Metrics: top-1 accuracy mAP mAP50 mAP75
Baselines: CE Focal Loss BBN LogitAdjust Mixup RISDA CMO RIDE KCL TSC BCL PaCo
Builds on: SupCon PaCo GLMC submodular information measures
Closest papers by meaning
- I-Con: A Unifying Framework for Representation Learning (ICLR 2025 · similarity 0.86)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · similarity 0.83)
- A Tale of Two Classes: Adapting Supervised Contrastive Learning to Binary Imbalanced Datasets (CVPR 2025 · similarity 0.83)
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · similarity 0.82)
- A Unified Loss for Handling Inter-Class and Intra-Class Imbalance in Medical Image Segmentation (AAAI 2025 · similarity 0.82)
- Stabilizing Sample Similarity in Representation via Mitigating Random Consistency (ICML 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.