Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Meta-Exploiting Frequency Prior for Cross-Domain Few-Shot 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 meta-learning framework for cross-domain few-shot learning that decomposes query images into FFT low- and high-frequency components and regularizes a ResNet-10 embedding with prediction-consistency (KL) and feature-reconstruction priors, adding no inference cost.
Problem
Meta-learning few-shot models overfit to source-domain task and semantic priors, degrading generalization when target domains differ (cross-domain few-shot learning).
Approach
Decompose each image via FFT into low- and high-frequency components processed by parallel EMA-updated branches, and regularize the main branch with a KL prediction-consistency prior and a feature-reconstruction prior so that full-band and frequency-band views agree; only the main branch is kept at test time.
Key results
- Direct inference reaches 46.85% (1-shot) and 63.77% (5-shot) averaged over eight target domains, beating LDP-net (46.34% / 62.60%).
- With query samples used for inference, average accuracy is 51.95% (1-shot) and 65.50% (5-shot), versus 50.85% / 64.10% for LDP-net†.
- Ablation on the meta-baseline shows average gains of 2.26% (1-shot) and 3.10% (5-shot) over the Meta-baseline; the method fails to beat all competitors on ChestX.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed frequency-prior meta-learning framework instead of LDP-net | Direct cross-domain FSL · 5-way 1-shot, 8-domain average | accuracy (%): 46.34 -> 46.85 | Helped |
| Use the proposed frequency-prior meta-learning framework instead of LDP-net | Direct cross-domain FSL · 5-way 5-shot, 8-domain average | accuracy (%): 62.60 -> 63.77 | Helped |
| Use the proposed method with query samples assisting inference instead of LDP-net† | Query-assisted cross-domain FSL · 5-way 1-shot, 8-domain average | accuracy (%): 50.85 -> 51.95 | Helped |
| Use the proposed method with query samples assisting inference instead of LDP-net† | Query-assisted cross-domain FSL · 5-way 5-shot, 8-domain average | accuracy (%): 64.10 -> 65.50 | Helped |
| Add the full frequency-prior method instead of the pretraining baseline | Cross-domain FSL · 5-way 1-shot, 8-domain average | accuracy (%): 51.37 -> 54.16 | Helped |
| Add the full frequency-prior method instead of the meta baseline | Cross-domain FSL · 5-way 5-shot, 8-domain average | accuracy (%): 71.77 -> 74.87 | Helped |
| Add the feature reconstruction prior (ablation) | Meta-baseline · 5-way 1-shot, 8-domain average | accuracy (%): 53.41 -> 54.16 | Helped |
| Add the feature reconstruction prior (ablation) | Meta-baseline · 5-way 5-shot, 8-domain average | accuracy (%): 73.41 -> 74.87 | Helped |
| Add the prediction consistency prior (ablation) | Meta-baseline · 5-way 1-shot, 8-domain average | accuracy (%): 51.90 -> 53.51 | Helped |
| Add the prediction consistency prior (ablation) | Meta-baseline · 5-way 5-shot, 8-domain average | accuracy (%): 71.77 -> 74.22 | Helped |
| Replace frequency priors with texture prior (ablation) | Frequency-prior method · 5-way 1-shot, 8-domain average | accuracy (%): 54.16 -> 52.32 | Hurt |
| Replace frequency priors with shape prior (ablation) | Frequency-prior method · 5-way 5-shot, 8-domain average | accuracy (%): 74.87 -> 70.61 | Hurt |
Surprising findings
- Triple-parameter baseline gains only modestly, supporting that improvements are not from added parameters.
- The method underperforms on the ChestX dataset, where the fixed FFT decomposition is argued to be suboptimal.
Limitations
- Fails to beat all competitors on the Chest dataset.
- Fixed FFT/wavelet decomposition may not be optimal; the authors plan learnable decomposition.
- Adds training-time overhead from decomposition and extra branches.
Experimental setup
Datasets: mini-ImageNet CUB Cars Places Plantae ChestX ISIC EuroSAT CropDisease
Benchmarks: CUB Cars Places Plantae ChestX ISIC EuroSAT CropDisease
Metrics: average classification accuracy (%) inference time per task
Baselines: MatchingNet RelationNet GNN FWT LRP ATA AFA LDP-net StyleAdv RDC Fine-tuning Meta-baseline
Builds on: prototypical networks meta-baseline LDP-net
Closest papers by meaning
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.90)
- HAP: Harmonized Amplitude Perturbation for Cross-Domain Few-Shot Learning (AAAI 2026 · similarity 0.89)
- Adaptive Cross-Modal Few-shot Learning (NeurIPS 2019 · similarity 0.88)
- Discriminative Sample-Guided and Parameter-Efficient Feature Space Adaptation for Cross-Domain Few-Shot Learning (CVPR 2024 · similarity 0.86)
- Flatten Long-Range Loss Landscapes for Cross-Domain Few-Shot Learning (CVPR 2024 · similarity 0.86)
- One Meta-tuned Transformer is What You Need for Few-shot Learning (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.