Research map / Label-efficient and robust learning / Recognition and long-tail
Long-tailed Recognition with Model Rebalancing
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 MORE, which decomposes each conv weight into a generalizable part and a low-rank minority part, trained with a discrepancy loss and sinusoidal schedule, to rebalance model capacity for long-tailed single- and multi-label recognition at no inference cost.
Problem
Long-tailed recognition, where models overfit head classes and generalize poorly to tail classes, especially in multi-label settings and CLIP finetuning where existing rebalancing, augmentation, and decoupling methods give inconsistent gains.
Approach
Decompose each weight matrix into a shared generalizable component plus a low-rank tail-specific component, and train the low-rank part with an l2 logit-discrepancy loss weighted by class priors and a sinusoidal schedule, so capacity is reallocated toward minority classes; the parts are merged at inference.
Key results
- On CIFAR-100-LT IF=10, ProCo+MORE reaches 65.9% top-1 vs 65.0% for ProCo; on IF=100 it reaches 52.9% vs 51.9%.
- On MIML from scratch, BCE+MORE improves overall mAP from 80.8 to 84.6.
- On ImageNet-LT, ProCo+MORE achieves 57.2% top-1 vs 56.3% for ProCo.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add MORE to LA | LA on CIFAR-100-LT IF=10 · CIFAR-100-LT IF=10 All | top-1 accuracy: 63.7 -> 64.8 | Helped |
| Add MORE to LA | LA on Places-LT · Places-LT Few | top-1 accuracy: 27.4 -> 30.6 | Helped |
| Add MORE to BCE | BCE on MIML, from scratch · MIML All | mAP: 80.8 -> 84.6 | Helped |
| Add MORE to ASL | ASL with CLIP finetuning on PASCAL-VOC · PASCAL-VOC Few | mAP: 82.2 -> 90.5 | Helped |
| Use sinusoidal rather than constant weighting (ablation) | MORE with LA on CIFAR-100-LT IF=10 · CIFAR-100-LT IF=10 | top-1 accuracy: 63.9 -> 64.8 | Helped |
| Use sinusoidal rather than cosine weighting (ablation) | MORE with LA on CIFAR-100-LT IF=10 · CIFAR-100-LT IF=10 | top-1 accuracy: 64.0 -> 64.8 | Helped |
| Use sinusoidal rather than constant weighting (ablation) | MORE with BCE on MIML · MIML | mAP: 82.3 -> 84.6 | Helped |
| Use cosine rather than sinusoidal weighting (ablation) | MORE across reported single-label and multi-label benchmarks · CIFAR-100-LT and multi-label datasets | performance | Hurt |
| Use LoRA under BCE instead of MORE (ablation) | BCE on PASCAL-VOC · PASCAL-VOC | mAP: 60.7 -> 58.4 | Hurt |
| Use parameter decomposition with BCE but without the MORE loss (ablation) | BCE on VOC and COCO · VOC and COCO | overall performance | No clear effect |
| Add the MORE loss to low-rank parameter decomposition (ablation) | BCE on VOC and COCO · VOC and COCO | overall performance | Helped |
| Replace the ℓ2 discrepancy with KL divergence (ablation) | MORE discrepancy metric · reported discrepancy-metric experiments | performance | Hurt |
Surprising findings
- Low-rank decomposition alone (MORE without the reallocation loss) performs comparably to the baseline, so the gains depend on the discrepancy loss.
- Under CLIP finetuning on MS-COCO, MORE slightly reduces Many-class mAP for BCE (52.9 to 50.4).
Limitations
- Decomposition is applied uniformly across layers with a fixed rank, without adaptive allocation.
- Validated only on visual recognition tasks; other modalities not tested.
- Tail classes share a single low-rank space, which may cause trade-offs among semantically different tail classes.
Experimental setup
Datasets: CIFAR-100-LT Places-LT ImageNet-LT MIML Pascal-VOC 2007 NUS-WIDE-SCENE MS-COCO
Benchmarks: CIFAR-100-LT Places-LT ImageNet-LT MIML Pascal-VOC 2007 NUS-WIDE-SCENE MS-COCO
Metrics: top-1 accuracy mAP Many/Medium/Few split accuracy
Baselines: CE CB BCL LA ProCo BCE Focal loss ASL RIDE BalPoE
Builds on: LoRA logit adjustment ProCo class-balanced loss
Closest papers by meaning
- Difficulty-aware Balancing Margin Loss for Long-tailed Recognition (AAAI 2025 · similarity 0.91)
- Exploring Weight Balancing on Long-Tailed Recognition Problem (ICLR 2024 · similarity 0.90)
- Long-Tailed Recognition via Information-Preservable Two-Stage Learning (NeurIPS 2025 · similarity 0.89)
- Long-Tailed Learning as Multi-Objective Optimization (AAAI 2024 · similarity 0.89)
- MLC-NC: Long-Tailed Multi-Label Image Classification Through the Lens of Neural Collapse (AAAI 2025 · similarity 0.88)
- Feature Fusion from Head to Tail for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.88)
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.