Research map / Label-efficient and robust learning / Recognition and long-tail
MLC-NC: Long-Tailed Multi-Label Image Classification Through the Lens of Neural Collapse
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 MLC-NC, a neural-collapse-inspired method for long-tailed multi-label image classification that uses fixed ETF label embeddings for class-wise feature extraction, collapse calibration, and a binarized fixed ETF classifier.
Problem
Long-tailed multi-label image classification, where imbalanced class distributions, head-tail label co-occurrence within instances, and label dependencies degrade learned representations and classifier bias.
Approach
Induce neural collapse properties in LT-MLC by extracting class-wise features via cross-attention with a fixed simplex ETF label embedding, aligning them to the embeddings, calibrating within-class variance in a projected low-dimensional space with a contrastive loss, and classifying with a fixed binarized ETF classifier over concatenated class features.
Key results
- On COCO-LT, MLC-NC reaches 60.52 total mAP vs 56.90 for URS, and 64.21 tail mAP vs 57.08 for MFM.
- On VOC-LT, MLC-NC reaches 84.37 total mAP vs 81.44 for URS and 90.31 tail mAP.
- On imbalanced VG200, MLC-NC reaches 36.87 total mAP vs 30.83 for the best DB baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MLC-NC instead of URS | ResNet50 multi-label classifier · COCO-LT | mAP (%) — Total: 56.90 -> 60.52 | Helped |
| add MLC-NC instead of URS | ResNet50 multi-label classifier · COCO-LT | mAP (%) — Head: 54.13 -> 49.69 | Hurt |
| add MLC-NC instead of URS | ResNet50 multi-label classifier · COCO-LT | mAP (%) — Medium: 60.59 -> 64.94 | Helped |
| add MLC-NC instead of MFM | ResNet50 multi-label classifier · COCO-LT | mAP (%) — Tail: 57.08 -> 64.21 | Helped |
| add MLC-NC instead of URS | ResNet50 multi-label classifier · VOC-LT | mAP (%) — Total: 81.44 -> 84.37 | Helped |
| add MLC-NC instead of MFM | ResNet50 multi-label classifier · VOC-LT | mAP (%) — Tail: 85.83 -> 90.31 | Helped |
| add MLC-NC instead of DB | ResNet50 multi-label classifier · VG200 | mAP (%) — Total: 30.83 -> 36.87 | Helped |
| add ETF label embedding guided feature learning (ablation) | MLC-NC · VOC-LT | mAP (%) — Total: 82.13 -> 84.37 | Helped |
| add feature projection and collapse calibration (ablation) | MLC-NC · VOC-LT | mAP (%) — Total: 83.43 -> 84.37 | Helped |
| add binarized fixed ETF classifier and concatenate features (ablation) | MLC-NC · VOC-LT | mAP (%) — Total: 79.59 -> 84.37 | Helped |
| replace fixed ETF label embedding with randomly initialized trainable label embedding (ablation) | MLC-NC EGFL · VOC-LT | mAP (%) — Total: 84.37 -> 83.92 | Hurt |
| replace fixed ETF label embedding with fixed GloVe label embedding (ablation) | MLC-NC EGFL · VOC-LT | mAP (%) — Total: 84.37 -> 83.74 | Hurt |
Surprising findings
- Projecting features to a lower dimension (20) yields better mAP than the original 768-dimensional features, and overall mAP tends to decline as dimension increases.
Limitations
- Performance on head classes is lower than some baselines (e.g., URS, DB) on COCO-LT and VOC-LT.
- Gains rely on a fixed ETF label embedding; trainable and GloVe embeddings perform slightly worse, indicating sensitivity to label embedding choice.
Experimental setup
Datasets: COCO-LT VOC-LT VG200
Benchmarks: COCO-LT VOC-LT VG200
Metrics: mAP head/medium/tail mAP global mAP
Baselines: ML-GCN Focal Loss ASL ERM RS RW OLTR LDAM CB Focal BBN DB DB-Focal
Builds on: neural collapse fixed ETF classifier (Yang et al. 2022) DB loss ASL TW loss (Kobayashi 2023)
Closest papers by meaning
- Neural Collapse in Multi-label Learning with Pick-all-label Loss (ICML 2024 · similarity 0.89)
- Long-tailed Recognition with Model Rebalancing (NeurIPS 2025 · similarity 0.88)
- Taming the Sigmoid Bottleneck: Provably Argmaxable Sparse Multi-Label Classification (AAAI 2024 · similarity 0.87)
- Difficulty-aware Balancing Margin Loss for Long-tailed Recognition (AAAI 2025 · similarity 0.87)
- MaxSup: Overcoming Representation Collapse in Label Smoothing (NeurIPS 2025 · similarity 0.86)
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 2024 · similarity 0.86)
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.