Research map / Label-efficient and robust learning / Recognition and long-tail
Boundary Matters: A Bi-Level Active Finetuning Method
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 BiLAF, a one-shot active finetuning method that pairs ActiveFT-style core sample selection with an unsupervised density-based denoising step and a boundary score to pick decision-boundary samples from a pretrained DeiT-Small feature space.
Problem
Selecting a limited annotation budget for supervised finetuning of pretrained vision models; existing active learning methods are biased under batch selection and the prior active finetuning method (ActiveFT) emphasizes global distribution coverage while neglecting decision boundary samples.
Approach
A bi-level scheme: first pick pseudo-class centers (via ActiveFT or K-Means), then denoise each pseudo-class with iterative density-based clustering and select boundary samples in one shot using a label-free boundary score with an opponent penalty and iterative removal.
Key results
- On CIFAR100 BiLAF outperforms ActiveFT by roughly 3 points (e.g., 43.5% vs 40.7% at 2%, 62.8% vs 54.6% at 5%), and on ImageNet it improves about 1 point (e.g., 50.8% vs 50.1% at 1%).
- On CIFAR10 at 0.5% budget BiLAF (81.0%) is below ActiveFT (85.0%), but at 1% and 2% BiLAF reaches 89.2% and 92.5% vs ActiveFT's 88.2% and 90.1%.
- Selection time on CIFAR100 is comparable to ActiveFT (e.g., 19.2s vs 21.9s at 5%) and far below CoreSet and VAAL, which take hours.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BiLAF instead of ActiveFT for sample selection | DeiT-Small (DINO) finetuning · CIFAR10 | Top-1 accuracy: 85.0±0.4 -> 81.0±1.2; 88.2±0.4 -> 89.2±0.6; 90.1±0.2 -> 92.5±0.4 | Mixed |
| use BiLAF instead of ActiveFT for sample selection | DeiT-Small (DINO) finetuning · CIFAR100 | Top-1 accuracy: 26.1±2.6 -> 31.8±1.6; 40.7±0.9 -> 43.5±0.8; 54.6±2.3 -> 62.8±1.2; 71.0±0.5 -> 73.7±0.5 | Helped |
| use BiLAF instead of ActiveFT for sample selection | DeiT-Small (DINO) finetuning · ImageNet-1k | Top-1 accuracy: 50.1±0.3 -> 50.8±0.4; 55.8±0.3 -> 56.9±0.3; 65.3±0.1 -> 66.2±0.2 | Helped |
| replace IDC denoising with DB denoising (ablation) | BiLAF · CIFAR100 2% | Top-1 accuracy: 43.48 -> 41.06 (−2.42) | Hurt |
| replace IDC denoising with distance-guide denoising (ablation) | BiLAF · CIFAR100 2% | Top-1 accuracy: 43.48 -> 36.42 (−7.06) | Hurt |
| remove the denoising process (ablation) | BiLAF · CIFAR100 2% | Top-1 accuracy: 43.48 -> 36.78 (−6.70) | Hurt |
| replace Boundary Score with the basic distance metric (ablation) | BiLAF · CIFAR100 2% | Top-1 accuracy: 43.48 -> 33.76 (−9.72) | Hurt |
| remove opponent penalty (ablation) | BiLAF · CIFAR100 | Top-1 accuracy: 43.48 -> 43.09 (−0.39); 31.82 -> 32.13 (+0.31) | Mixed |
| replace iterative selection and removal with one-shot selection (ablation) | BiLAF · CIFAR100 | Top-1 accuracy: 43.48 -> 42.13 (−1.35); 62.75 -> 60.24 (−2.51); 73.67 -> 69.66 (−4.01) | Hurt |
| vary the number of core samples (ablation) | BiLAF · CIFAR100 1% | Top-1 accuracy: 21.51; 31.82; 28.37; 27.24 | Mixed |
| replace ActiveFT with Random, FDS, or K-Means for core selection (ablation) | BiLAF · CIFAR100 | Top-1 accuracy: 25.58; 20.69; 28.80; 31.82 | Mixed |
Surprising findings
- K-Means as core selection method outperforms ActiveFT at the 10% budget on CIFAR100 (74.27 vs 73.67).
- Removing the opponent penalty slightly improves performance at the 1% budget (32.13 vs 31.82).
Limitations
- Performance is weaker at very low budgets (e.g., CIFAR10 at 0.5%) where boundary samples add instability.
- Hyperparameters such as removal ratio and core number need tuning; the removal rate can remove minority samples in long-tailed data.
- Less pronounced gains on object detection where global features may not capture multi-instance structure.
Experimental setup
Datasets: CIFAR10 CIFAR100 ImageNet-1k PASCAL VOC ADE20K CIFAR100LT CUB-200-2011
Benchmarks: CIFAR10 CIFAR100 ImageNet-1k PASCAL VOC ADE20K CIFAR100LT CUB-200-2011
Metrics: Top-1 accuracy mAP mIoU entropy ProbDiff running time
Baselines: Random FDS K-Means CoreSet VAAL LearnLoss TA-VAAL ALFA-Mix ActiveFT
Builds on: ActiveFT DBSCAN CoreSet DINO
Closest papers by meaning
- ActiveDC: Distribution Calibration for Active Finetuning (CVPR 2024 · similarity 0.88)
- Uncertainty-aware Fine-tuning of Segmentation Foundation Models (NeurIPS 2024 · similarity 0.85)
- Diffusion-Driven Two-Stage Active Learning for Low-Budget Semantic Segmentation (NeurIPS 2025 · similarity 0.85)
- A Cross-Domain Benchmark for Active Learning (NeurIPS 2024 · similarity 0.84)
- Patch-Aware Sample Selection for Efficient Masked Image Modeling (AAAI 2024 · similarity 0.84)
- Sketchy Moment Matching: Toward Fast and Provable Data Selection for Finetuning (NeurIPS 2024 · similarity 0.84)
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.