Research map / Label-efficient and robust learning / Recognition and long-tail
Class and Attribute-Aware Logit Adjustment for Generalized Long-Tail 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 a class-attribute aware logit-adjusted (CALA) loss that adds a class-conditional probability density ratio term to class-prior logit adjustment, with a K-neighborhood heuristic and a meta-learned adjustment network to estimate the ratio for generalized long-tail learning.
Problem
Long-tail learning under both class imbalance and attribute imbalance (generalized long-tail), where the training and test class-conditional densities differ, so existing logit adjustment based only on class priors is insufficient.
Approach
Derives a logit-adjusted loss that includes both the class prior and the class-conditional density ratio between training and test data, estimates that ratio from neighborhood information (Heuristic-CALA) or via a meta-learned adjustment network over training-dynamics features (Meta-CALA).
Key results
- Meta-CALA reaches 84.79% on CIFAR10-LT (100:1) and 52.34% on CIFAR100-LT (100:1), exceeding the best prior baselines.
- Heuristic-CALA reaches 54.13% accuracy on ImageNet-GLT CLT protocol versus 46.74% for LDAM, and 44.71% on the GLT protocol.
- On Waterbirds, Heuristic-CALA reaches 94.3% average and 91.8% worst-group accuracy, outperforming the second-best baselines by 1.65% and 1.15%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Meta-CALA instead of the strongest listed baseline | CIFAR10-LT · CIFAR10-LT, 100:1 | accuracy: 82.10% -> 84.79% | Helped |
| Use Meta-CALA instead of the strongest listed baseline | CIFAR10-LT · CIFAR10-LT, 10:1 | accuracy: 90.80% -> 92.47% | Helped |
| Use Meta-CALA instead of the strongest listed baseline | CIFAR100-LT · CIFAR100-LT, 100:1 | accuracy: 48.71% -> 52.34% | Helped |
| Use Meta-CALA instead of the strongest listed baseline | CIFAR100-LT · CIFAR100-LT, 10:1 | accuracy: 61.73% -> 65.51% | Helped |
| Use Meta-CALA instead of the strongest listed baseline | iNat 2018 · iNat 2018 | accuracy: 72.01% -> 74.05% | Helped |
| Use Meta-CALA instead of the strongest listed baseline | Places-LT · Places-LT | accuracy: 42.64% -> 43.97% | Helped |
| Use Heuristic-CALA instead of the strongest listed baseline | Waterbirds · Waterbirds | average accuracy: 92.4% -> 94.3% | Helped |
| Use Heuristic-CALA instead of the strongest listed baseline | Waterbirds · Waterbirds | worst-group accuracy: 90.6% -> 91.8% | Helped |
| Use Heuristic-CALA instead of the strongest listed baseline | CMNIST · CMNIST | average accuracy: 78.1% -> 79.5% | Helped |
| Use Heuristic-CALA instead of the strongest listed baseline | CMNIST · CMNIST | worst-group accuracy: 75.9% -> 77.0% | Helped |
| Add the class prior term u(y) (ablation) | CALA · ImageNet-GLT GLT | accuracy: 42.51% -> 44.71% | Helped |
| Add the CCPD ratio term v(x, y) (ablation) | CALA · ImageNet-GLT GLT | accuracy: 37.80% -> 44.71% | Helped |
Surprising findings
- Removing the CCPD ratio term v(x,y) hurts GLT accuracy more than removing the class-prior term u(y), with ImageNet-GLT GLT accuracy dropping from 44.71% to 37.80%.
Limitations
- Meta-CALA requires an additional balanced metadata set and increases training time complexity.
- Hyperparameters (tau1, tau2, neighborhood size K) are tuned per benchmark and sensitivity is only shown on CIFAR-LT.
Experimental setup
Datasets: CIFAR10-LT CIFAR100-LT Places-LT iNat 2018 Waterbirds CMNIST CelebA ImageNet-GLT MSCOCO-GLT
Benchmarks: CIFAR10-LT CIFAR100-LT iNat 2018 Places-LT Waterbirds CMNIST ImageNet-GLT MSCOCO-GLT
Metrics: accuracy precision average accuracy worst-group accuracy
Baselines: Class-Balanced CE Class-Balanced Focal LDAM-DRW De-confound-TDE LA MiSLAS LADE GLC ALA BKD CSA Meta-Weight-Net
Builds on: logit adjustment (LA) LDAM ALA MixUp Meta-Weight-Net invariant feature learning (IFL)
Closest papers by meaning
- Improved Balanced Classification with Theoretically Grounded Loss Functions (NeurIPS 2025 · similarity 0.87)
- Long-Tailed Learning as Multi-Objective Optimization (AAAI 2024 · similarity 0.86)
- Balancing the Scales: A Theoretical and Algorithmic Framework for Learning from Imbalanced Data (ICML 2025 · similarity 0.85)
- Rethinking the Bias of Foundation Model under Long-tailed Distribution (ICML 2025 · similarity 0.84)
- Balancing Model Efficiency and Performance: Adaptive Pruner for Long-tailed Data (ICML 2025 · similarity 0.84)
- Robust Logit Adjustment for Learning with Long-Tailed Noisy Data (AAAI 2025 · 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.