Research map / Label-efficient and robust learning / Noisy labels
Dual-Level Curriculum Meta-Learning for Noisy Few-Shot Learning Tasks
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 DCML, a dual-level curriculum meta-learning method that samples classes via class-pair boundary-complexity metrics and clean examples via an under-trained proxy model, evaluated on noisy few-shot benchmarks.
Problem
Few-shot learning is highly vulnerable to label noise in both meta-training and meta-test tasks, and prior noisy-FSL methods only handle example-level noise and one side of the noise problem.
Approach
A dual-level curriculum: class-level sampling uses CP-metrics (historical mean and variance of incorrect-class probabilities across checkpoints) to separate easy, similar, and noisy class pairs, while example-level sampling keeps low-loss examples using a fixed, under-trained earlier checkpoint as a proxy; the paper also adds convergence and hierarchical PAC-Bayes bounds.
Key results
- On 5-way 1-shot with asymmetric noise on miniImageNet, DCML reaches 0.5315 vs 0.5201 for IDEAL and 0.4612 for MAML.
- With 40% symmetric label noise injected at meta-test, DCML gets 0.3573 on miniImageNet 1-shot, the top among compared methods.
- On real-world noisy Food101 (5-way 1-shot), DCML reaches 0.5165 vs 0.4844 for RNNP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use dual-level class-example curriculum sampling instead of the strongest reported baseline | Noisy few-shot learning · miniImageNet, FC100, and Omniglot with asymmetric, symmetric, and background noise; 5-way 1-shot | accuracy: 0.5315 vs 0.5100; 0.4671 vs 0.4410; 0.9677 vs 0.9643; 0.4627 vs 0.3852; 0.4822 vs 0.4606; 0.9543 vs 0.9347; 0.5193 vs 0. | Helped |
| Use dual-level class-example curriculum sampling instead of the strongest reported baseline | Noisy few-shot learning · CIFAR-100N, Food101FS, and miniWV; 5-way 1-shot | accuracy: 0.6059 vs 0.5928; 0.5165 vs 0.4844; 0.3986 vs 0.3815 | Helped |
| Add class-level curriculum learning (C-CL) to a model without curriculum learning (ablation) | DCML · CIFAR-100N, Food101, and miniWV, 5-way 5-shot | accuracy | Helped |
| Add example-level curriculum learning (E-CL) to a model without curriculum learning (ablation) | DCML · CIFAR-100N, Food101, and miniWV, 5-way 5-shot | accuracy | Helped |
| Combine C-CL and E-CL instead of using either level alone (ablation) | DCML · CIFAR-100N, Food101, and miniWV, 5-way 5-shot | accuracy | Helped |
| Add 40% symmetric label noise to the support set during meta-test | Few-shot classifiers · miniImageNet, FC100, Omniglot, CIFAR-100N, Food101FS, and miniWV; 5-way 1-shot | accuracy: up to 10% | Hurt |
| Increase the synthetic label-noise ratio | Few-shot learning methods · miniImageNet | accuracy: 10%, 20%, 30%, 40%, 50% | Hurt |
| Increase the mixture-noise ratio from 30% to 50% | DCML · miniImageNet | accuracy: 30% to 50%; 0.04% | Hurt |
Surprising findings
- Performance of baseline few-shot methods drops by up to about 10% when symmetric noise is added to the support set at meta-test time.
Limitations
- Class-pair thresholds depend on estimated noise ratio, which may need a clean subset to estimate.
- Theoretical bounds rely on assumptions such as stable performance across subjects that are not guaranteed in general.
Experimental setup
Datasets: miniImageNet FC100 Omniglot miniWV Food101 CIFAR-100N
Benchmarks: miniImageNet FC100 Omniglot miniWV Food101 CIFAR-100N
Metrics: 5-way 1-shot accuracy 5-way 5-shot accuracy
Baselines: MAML MAML+spld MAML+focal MAML+dih CMAML RNNP RapNets TraNFS CT IDEAL
Builds on: MAML self-paced learning PAC-Bayesian meta-learning
Closest papers by meaning
- A Nested Bi-level Optimization Framework for Robust Few Shot Learning (AAAI 2022 · similarity 0.87)
- Bridging the Gap Between Practice and PAC-Bayes Theory in Few-Shot Meta-Learning (NeurIPS 2021 · similarity 0.87)
- Robust Curriculum Learning: from clean label detection to noisy label self-correction (ICLR 2021 · similarity 0.86)
- SuperLoss: A Generic Loss for Robust Curriculum Learning (NeurIPS 2020 · similarity 0.86)
- Robust Meta-learning with Sampling Noise and Label Noise via Eigen-Reptile (ICML 2022 · similarity 0.86)
- Improving Task-Specific Generalization in Few-Shot Learning via Adaptive Vicinal Risk Minimization (NeurIPS 2022 · 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.