Research map / Label-efficient and robust learning / Recognition and long-tail
Multi-Head Encoding for Extreme Label Classification
Daojun Liang, Haixia Zhang, Dongfeng Yuan, Minggao Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace vanilla classifier with MHS | ResNet-101 classifier · VOC2007 label powerset | accuracy: 79.82 -> 80.34 | Improved |
| replace vanilla classifier with MHS | ResNet-101 classifier · COCO2014 label powerset | accuracy: 46.62 -> 47.25 | Improved |
| use MHE instead of AttentionXML HLT | XLC classifier · COCO2014 label powerset | accuracy: 46.06 -> 47.21 | Improved |
| MHC without label clustering instead of AttentionXML | Single model · Wiki-500K | P@1: 75.01 -> 78.38 | Improved |
| longer main head {172;23} instead of {86;46} (ablation) | MHC · Eurlex-4K | P@1: 85.54 -> 86.00 | Improved |
| longer main head {310;43} instead of {155;86} (ablation) | MHC · AmazonCat-13K | P@1: 95.83 -> 95.94 | Improved |
| add MHS pretraining instead of ArcFace-Full | ResNet-18 · CASIA IJB-C | TAR: 94.31 -> 95.06 | Improved |
| add MHS pretraining instead of ArcFace-Full | ResNet-101 · MS1MV2 IJB-C | TAR: 96.20 -> 95.85 | Hurt |
| add label clustering instead of random decomposition (ablation) | small model SimpleConv-4 · generalization | accuracy: about 4% | Improved |
| add label clustering instead of random decomposition (ablation) | overfit ResNet-18 · generalization | accuracy: the gap between LRD and LC vanishes | No clear effect |
| use MHP instead of one-hot encoding classifier | classifier · speed at 1 million labels | runtime: 17x faster than OHE | Improved |
| use MHS instead of OHE | classifier · speed at 1 million labels | runtime: 8x faster than OHE | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- AttentionXML — AttentionXML: Label Tree-based Attention-Aware Deep Model for High-Performance Extreme Multi-Label Text Classification (NeurIPS 2019) · better on 4 of 4 shared results
- LightXML — LightXML: Transformer with Dynamic Negative Sampling for High-Performance Extreme Multi-label Text Classification (AAAI 2021) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Taming the Sigmoid Bottleneck: Provably Argmaxable Sparse Multi-Label Classification (AAAI 2024 · similarity 0.87)
- Self-Paced Unified Representation Learning for Hierarchical Multi-Label Classification (AAAI 2024 · similarity 0.84)
- Long-Tailed Partial Label Learning by Head Classifier and Tail Classifier Cooperation (AAAI 2024 · similarity 0.84)
- MLC-NC: Long-Tailed Multi-Label Image Classification Through the Lens of Neural Collapse (AAAI 2025 · similarity 0.84)
- ELMO : Efficiency via Low-precision and Peak Memory Optimization in Large Output Spaces (ICML 2025 · similarity 0.84)
- Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning (AAAI 2024 · similarity 0.84)
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.
