Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Context-Aware 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 CAML, a visual meta-learner that recasts few-shot image classification as non-causal sequence modeling over a frozen CLIP encoder with fixed ELMES label embeddings, pre-trained once and evaluated without meta-training or inference-time fine-tuning.
Problem
Visual meta-learners cannot learn new object classes at inference time like LLM in-context learning; existing methods either perform poorly in cross-domain settings or require meta-training or fine-tuning on the support set.
Approach
Treat n-way k-shot classification as non-causal sequence modeling over (image, label) support pairs plus an unlabeled query, using a frozen pre-trained image encoder and a fixed Equal Length and Maximally Equiangular Set (ELMES) label encoder, so the model learns new classes in context without fine-tuning.
Key results
- CAML without meta-training or fine-tuning reaches 96.2% 5w-1s on mini-ImageNet, above P>M>F's 95.3%.
- CAML matches or exceeds P>M>F on 8 of 11 benchmarks, per the paper's claim.
- CAML underperforms P>M>F on CIFAR-fs, Aircraft and ChestX; on Aircraft a learnable class embedding variant reaches 66.3% 5w-1s.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CAML rather than the strongest universal baseline | Universal few-shot image classification · CIFAR-fs 5w-1s | accuracy: 62.9$\pm.2$ -> 70.8$\mathbf{\pm.2}$ | Helped |
| Use CAML rather than the strongest universal baseline | Universal few-shot image classification · MiniImageNet 5w-1s | accuracy: 95.6$\pm.1$ -> 96.2$\mathbf{\pm.1}$ | Helped |
| Use CAML rather than the strongest universal baseline | Universal few-shot image classification · Pascal + Paintings 5w-1s | accuracy: 62.6$\pm.2$ -> 63.8$\mathbf{\pm.2}$ | Helped |
| Use CAML rather than the strongest universal baseline | Universal few-shot image classification · Paintings 5w-1s | accuracy: 51.9$\mathbf{\pm.2}$ -> 51.1$\pm.2$ | Hurt |
| Use CAML rather than the strongest universal baseline | Universal few-shot image classification · Pascal 5w-1s | accuracy: 81.7$\pm.2$ -> 82.6$\pm.2$ | Helped |
| Use CAML rather than the strongest universal baseline | Universal few-shot image classification · meta-iNat 5w-1s | accuracy: 86.3$\pm.2$ -> 91.2$\mathbf{\pm.2}$ | Helped |
| Use CAML rather than the strongest universal baseline | Universal few-shot image classification · tiered meta-iNat 5w-1s | accuracy: 77.3$\pm.2$ -> 81.9$\mathbf{\pm.2}$ | Helped |
| Use CAML rather than the strongest universal baseline | Universal few-shot image classification · ChestX 5w-1s | accuracy: 23.0$\mathbf{\pm.1}$ -> 21.5$\pm.1$ | Hurt |
| Use CAML rather than the strongest universal baseline | Universal few-shot image classification · CUB 5w-1s | accuracy: 88.3$\pm.2$ -> 91.8$\mathbf{\pm.2}$ | Helped |
| Use CAML rather than the strongest universal baseline | Universal few-shot image classification · tiered-ImageNet 5w-1s | accuracy: 94.6$\pm.1$ -> 95.4$\mathbf{\pm.1}$ | Helped |
| Use CAML rather than the strongest universal baseline | Universal few-shot image classification · Aircraft 5w-1s | accuracy: 64.8$\mathbf{\pm.2}$ -> 63.3$\pm.3$ | Hurt |
| Replace fixed ELMES class embeddings with learnable class embeddings (ablation) | CAML · Few-shot benchmarks | accuracy: 85.5$\mathbf{\pm.1}$ -> 85.9$\mathbf{\pm.1}$; 63.3$\pm.3$ -> 66.3$\mathbf{\pm.2}$ | Mixed |
Surprising findings
- Frozen CLIP embeddings are antagonistic for some texture-classification queries, where the query sits closest to the wrong support class.
- Fine-tuning the CLIP backbone during pre-training hurts ProtoNet and MetaOpt on out-of-domain tasks such as Aircraft, CUB and Paintings.
Limitations
- Lags in-domain state of the art on highly out-of-distribution images (chest x-rays) and on low-resolution CIFAR images.
- Performance on fine-grained Aircraft depends on the CLIP embedding space; a ViT-Huge Laion-2b encoder substantially improves it.
Experimental setup
Datasets: ImageNet-1k Fungi MSCOCO WikiArt mini-ImageNet tiered-ImageNet CIFAR-fs Pascal VOC CUB Aircraft meta-iNat tiered meta-iNat
Benchmarks: mini-ImageNet tiered-ImageNet CIFAR-fs Pascal VOC CUB Aircraft meta-iNat tiered meta-iNat ChestX Paintings Pascal+Paintings
Metrics: mean accuracy standard error
Baselines: ProtoNet MetaOpt MetaQDA SNAIL GPICL P>M>F
Builds on: Prototypical Networks SNAIL GPICL P>M>F MetaQDA
Closest papers by meaning
- Unsupervised Meta-Learning via In-Context Learning (ICLR 2025 · similarity 0.89)
- Beyond Induction Heads: In-Context Meta Learning Induces Multi-Phase Circuit Emergence (ICML 2025 · similarity 0.87)
- Provable Meta-Learning with Low-Rank Adaptations (NeurIPS 2025 · similarity 0.87)
- Activating Visual Context and Commonsense Reasoning Through Masked Prediction in VLMs (AAAI 2026 · similarity 0.87)
- Link-Context Learning for Multimodal LLMs (CVPR 2024 · similarity 0.86)
- Any-Way Meta Learning (AAAI 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.