Research map / Label-efficient and robust learning / Recognition and long-tail
Harnessing Hierarchical Label Distribution Variations in Test Agnostic Long-tail Recognition
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 DirMixE, a mixture-of-experts method for test-agnostic long-tail recognition that models unknown test label distributions as a Dirichlet mixture meta-distribution and optimizes mean plus semi-variance of logit-adjusted loss.
Problem
Long-tail recognition when test label distributions are unknown and arbitrarily imbalanced; prior MoE methods only target a few fixed test distributions and ignore local variations.
Approach
Model test label distributions hierarchically as a mixture of Dirichlet distributions (global variation across components, local variation within each), assign one expert per component, and train with a Monte Carlo estimate of mean plus semi-variance of the logit-adjusted loss over sampled test distributions.
Key results
- On CIFAR-100-LT (our setting), DirMixE reaches 53.34 mean accuracy vs. 51.77 for SADE and 48.17 for BalPoE.
- On ImageNet-LT (our setting), DirMixE achieves 61.50 mean accuracy vs. 60.71 for SADE.
- On CIFAR-10-LT (our setting), DirMixE gets 87.57 mean vs. 87.46 for SADE, with gains concentrated on backward-LT distributions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DirMixE instead of SADE | CIFAR-10-LT · CIFAR-10-LT (Ours Setting) | accuracy (Mean) | Helped |
| Use DirMixE instead of SADE | CIFAR-100-LT · CIFAR-100-LT (Ours Setting) | accuracy (Mean): 2.4 to 4.1 | Helped |
| Use DirMixE instead of SADE or BalPoE | ImageNet-LT · ImageNet-LT (Ours Setting) | accuracy (Mean) | Helped |
| Use DirMixE instead of SADE or BalPoE | CIFAR-100-LT · CIFAR-100-LT (Ours Setting) | accuracy | Helped |
| Model unseen test label distributions with a Dirichlet mixture rather than fixed distributions | DirMixE · Test-agnostic long-tail recognition | generalization error | Helped |
| Penalize semi-variance rather than full variance | DirMixE · CIFAR-10 and CIFAR-100 | semi-variance/variance ratio: ρ = 0.503, 0.509 | No clear effect |
Surprising findings
- On uniform test distributions DirMixE is comparable to or slightly behind SADE and BalPoE, with gains mainly on backward-LT distributions.
Limitations
- Gains over SADE and BalPoE are small on CIFAR-10 and uniform/forward test distributions.
- Theoretical semi-variance assumption is validated only empirically on CIFAR datasets.
Experimental setup
Datasets: CIFAR-10-LT CIFAR-100-LT ImageNet-LT iNaturalist
Benchmarks: CIFAR-10-LT CIFAR-100-LT ImageNet-LT iNaturalist
Metrics: top-1 accuracy mean accuracy across test label distributions standard deviation across test distributions
Baselines: LDAM LA VS LADE DDC RIDE SADE BalPoE
Builds on: SADE BalPoE logit adjustment RIDE
Closest papers by meaning
- Learning Label Shift Correction for Test-Agnostic Long-Tailed Recognition (ICML 2024 · similarity 0.91)
- DiffuLT: Diffusion for Long-tail Recognition Without External Knowledge (NeurIPS 2024 · similarity 0.88)
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 2024 · similarity 0.87)
- Long-tailed Recognition with Model Rebalancing (NeurIPS 2025 · similarity 0.87)
- LTGC: Long-tail Recognition via Leveraging LLMs-driven Generated Content (CVPR 2024 · similarity 0.86)
- Rectifying Soft-Label Entangled Bias in Long-Tailed Dataset Distillation (NeurIPS 2025 · similarity 0.85)
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.