Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Any-Way Meta 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 "any-way" meta-learning that samples task cardinality per episode and exploits label equivalence via multiple random numeric label assignments over a shared output pool, plus an auxiliary semantic classifier with mixup, applied to MAML and ProtoNet.
Problem
Meta-learning models are fixed to a training cardinality (number of ways) and degrade sharply on tasks with different numbers of classes unseen during training.
Approach
Exploit label equivalence from stochastic numeric label assignment: sample the way per episode, map its numeric labels to several non-overlapping subsets of a larger output layer (ensembling logits), and add an auxiliary semantic classifier trained with cross-entropy on true class labels to inject semantic information.
Key results
- a-MAML (5-shot) reaches 79.06% on miniImageNet 3-way versus 76.64% for f-MAML (5-shot), and 81.55% vs 79.54% on Cars 3-way.
- Training a fixed 10-way 1-shot f-MAML on miniImageNet gives 31.91% while a-MAML gets 31.34%, showing comparable rather than superior results at 10-way.
- Ensembling more label assignments improves 5-way miniImageNet 1-shot accuracy from 65% (1 assignment) to about 66.8-67% (18 assignments in the table).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace fixed-way training with any-way training | MAML · miniImageNet, 3-way 1-shot | accuracy: 59.56(1.48) -> 63.15 (0.52) | Helped |
| Replace fixed-way training with any-way training | MAML · miniImageNet, 10-way 1-shot | accuracy: 31.91 (0.50) -> 31.34(0.28) | Hurt |
| Add semantic classifier to any-way MAML (ablation) | a-MAML · miniImageNet, 3-way 1-shot | accuracy: 63.15 (0.52) -> 64.25 (1.42) | Helped |
| Add semantic classifier to any-way MAML (ablation) | a-MAML · miniImageNet, 5-way 5-shot | accuracy: 66.73(0.42) -> 66.91(0.10) | No clear effect |
| Add semantic classifier to any-way MAML (ablation) | a-MAML · Cars, 3-way 5-shot | accuracy: 81.55(0.49) -> 81.65(0.19) | Helped |
| Add semantic classifier to any-way MAML (ablation) | a-MAML · TieredImageNet, 5-way 5-shot | accuracy: 68.10(0.63) -> 67.40(1.29) | Hurt |
| Add semantic classifier to a-MAML (ablation) | a-MAML · miniImageNet to TieredImageNet, 3-way 1-shot | accuracy: 66.27(0.36) -> 66.26(0.70) | No clear effect |
| Increase the number of assignment sets from 1 to 18 (ablation) | a-MAML · miniImageNet, 5-way 1-shot | accuracy: 47.64 (0.53) -> 48.55 (0.34) | Helped |
Surprising findings
- Any-way training matches or beats fixed-way training on several settings, contrary to the common expectation that fixed-way specialization helps.
- Fixed-way 10-way 1-shot f-MAML often failed to train (corrupted runs), while any-way MAML did not.
Limitations
- Many tables are incomplete or inconsistent (empty ProtoNet and semantic rows, garbled text), weakening the reported evidence.
- Ensembling gains saturate and the 10-way results are roughly comparable to fixed-way rather than clearly superior.
- Semantic injection relies on labeled semantic classes, limiting use in purely unlabeled settings.
Experimental setup
Datasets: miniImageNet TieredImageNet Cars CUB
Benchmarks: miniImageNet TieredImageNet Cars CUB
Metrics: accuracy
Baselines: f-MAML (fixed-way MAML) ProtoNet
Builds on: MAML ProtoNet mixup
Closest papers by meaning
- Meta-Learning without Memorization (ICLR 2020 · similarity 0.88)
- Learning to Learn with Contrastive Meta-Objective (NeurIPS 2025 · similarity 0.88)
- Meta-Learning Requires Meta-Augmentation (NeurIPS 2020 · similarity 0.88)
- Task Groupings Regularization: Data-Free Meta-Learning with Heterogeneous Pre-trained Models (ICML 2024 · similarity 0.87)
- Context-Aware Meta-Learning (ICLR 2024 · similarity 0.86)
- Human-like Category Learning by Injecting Ecological Priors from Large Language Models into Neural Networks (ICML 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.