Research map / Architectures and efficiency / Neural architecture search
Neural Fine-Tuning Search for 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 Neural Fine-Tuning Search (NFTS), a supernet-based NAS that searches per-layer choices of freezing/fine-tuning and adapter insertion for few-shot adaptation, deferring final path selection to test episodes.
Problem
Few-shot recognition via fine-tuning requires choosing which layers to fine-tune and where to insert adapters; these choices are heuristic, and per-episode architecture selection risks overfitting to small support sets.
Approach
Searches a space of per-layer binary choices (adapter on/off, pre-trained vs fine-tuned weights) using a SPOS-style supernet and evolutionary search at training time, shortlists N diverse architectures, and selects among them per test episode using the support set.
Key results
- On Meta-Dataset single-domain with ResNet-18, NFTS reaches 75.2% average vs 73.3% for TSA.
- On Meta-Dataset multi-domain with ViT-S, NFTS reaches 83.4% average vs 83.1% for PMF.
- Meta-Album: NFTS dominates well-tuned baselines across 1-20 shots, with margins greater than 5% at 5-way 5-shot (as reported).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NFTS instead of TSA | ResNet-18 · Meta-Dataset single-domain | average accuracy: 73.3 -> 75.2 | Helped |
| Use NFTS instead of ETT | ViT-S · Meta-Dataset single-domain | average accuracy: 77.6 -> 79.2 | Helped |
| Use NFTS instead of TSA | ResNet-18 · Meta-Dataset multi-domain | average accuracy: 78.4 -> 80.7 | Helped |
| Use NFTS instead of PMF | ViT-S · Meta-Dataset multi-domain | average accuracy: 83.1 -> 83.4 | Helped |
| Add adapters at all layers to a frozen backbone (ablation) | ResNet-18 · Meta-Dataset single-domain | average accuracy: 67.8 -> 70.4 | Helped |
| Fine-tune all backbone layers without adapters (ablation) | ResNet-18 · Meta-Dataset single-domain | average accuracy: 67.8 -> 70.2 | Helped |
| Add adapters and fine-tune all layers (ablation) | ResNet-18 · Meta-Dataset single-domain | average accuracy: 67.8 -> 70.8 | Helped |
| Select among preselected architectures per episode instead of fixing the top-1 architecture (ablation) | NFTS, ResNet-18 · Meta-Dataset single-domain | average accuracy: 73.6 -> 75.2 | Helped |
| Add adapters at all layers to a frozen backbone (ablation) | ViT-S · Meta-Dataset single-domain | average accuracy: 71.8 -> 73.8 | Helped |
| Fine-tune all backbone layers without adapters (ablation) | ViT-S · Meta-Dataset single-domain | average accuracy: 71.8 -> 74.0 | Helped |
| Add adapters and fine-tune all layers (ablation) | ViT-S · Meta-Dataset single-domain | average accuracy: 71.8 -> 74.4 | Helped |
| Select among preselected architectures per episode instead of fixing the top-1 architecture (ablation) | NFTS, ViT-S · Meta-Dataset single-domain | average accuracy: 78.7 -> 79.2 | Helped |
Surprising findings
- Single fixed architecture found by search (NFTS-1) already beats hand-designed adaptation strategies, but deferring selection to test episodes (NFTS-N) adds further gains, especially for ResNet multi-domain.
Limitations
- Coarse search space limited to two binary decisions per layer and fixed adapter types (TSA for ResNet, ETT prefix for ViT).
- Test-time selection fine-tunes and evaluates on the same small support set, which risks overfitting.
- Test-time cost is about 3x that of a single adaptation.
Experimental setup
Datasets: Meta-Dataset Meta-Album
Benchmarks: Meta-Dataset Meta-Album
Metrics: mean accuracy over episodes
Baselines: FLUTE ProtoNet BOHB FO-MAML TSA PMF ETT CNAPS SCNAPS SUR URT URL
Builds on: SPOS TSA ETT Prototypical Networks URL
Closest papers by meaning
- Partial Is Better Than All: Revisiting Fine-tuning Strategy for Few-shot Learning (AAAI 2021 · similarity 0.91)
- Revisiting Neural Networks for Few-Shot Learning: A Zero-Cost NAS Perspective (ICML 2025 · similarity 0.90)
- Efficient Few-Shot Neural Architecture Search by Counting the Number of Nonlinear Functions (AAAI 2025 · similarity 0.88)
- Strong Baselines for Parameter-Efficient Few-Shot Fine-Tuning (AAAI 2024 · similarity 0.88)
- einspace: Searching for Neural Architectures from Fundamental Operations (NeurIPS 2024 · similarity 0.87)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.87)
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.