Research map / Label-efficient and robust learning / Recognition and long-tail
ActiveDC: Distribution Calibration for Active Finetuning
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 ActiveDC, which selects samples for active finetuning via continuous distribution matching (ActiveFT) and then calibrates the selected set's distribution by generating and pseudo-labeling real unlabeled features using clustered, Tukey-transformed statistics.
Problem
Active finetuning of pretrained vision models with very small labeled budgets yields subsets with biased distributions that cause overfitting and poor accuracy.
Approach
Augments the ActiveFT-selected labeled set by sampling Gaussian features from calibrated per-cluster statistics (Tukey power transform, K-Means pseudo-categories, shrinkage of means toward labeled centers), mapping them to nearest real unlabeled features, pseudo-labeling them, and filtering with EMD.
Key results
- On CIFAR10 at 0.1% sampling, ActiveDC reaches 61.3% top-1 vs 47.1% for ActiveFT.
- On CIFAR100 at 1% sampling, ActiveDC reaches 34.5% vs 26.1% for ActiveFT at 1%... (gains reported up to ~10%+ at low ratios).
- Distribution calibration adds only minutes of selection time (e.g., 4m30s vs 21.9s for ActiveFT at 5% on CIFAR100) while improving accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ActiveDC distribution calibration to ActiveFT selection | ActiveFT · CIFAR10 (0.1%, 0.2%, 0.5%, 1%, 2%) | Top-1 accuracy: 47.1 -> 61.3; 64.5 -> 73.1; 85.0 -> 87.3; 88.2 -> 88.9; 90.1 -> 90.3 | Helped |
| Add ActiveDC distribution calibration to ActiveFT selection | ActiveFT · CIFAR100 (1%, 2%, 5%, 10%) | Top-1 accuracy: 26.1 -> 34.5; 40.7 -> 54.6; 54.6 -> 71.9; 71.0 -> 74.3 | Helped |
| Add ActiveDC distribution calibration to ActiveFT selection | ActiveFT · ImageNet-1k (0.5%, 1%, 2%, 5%) | Top-1 accuracy: 36.8 -> 50.9; 50.1 -> 56.3; 54.2 -> 60.1; 65.3 -> 68.2 | Helped |
| Add the Distributed Calibration Module (ablation) | ActiveFT · CIFAR10 (0.1%, 0.2%) | Top-1 accuracy: 47.1 -> 59.5; 64.5 -> 70.6 | Helped |
| Set the generated-feature count to 2× rather than 1×, 3×, or 4× (ablation) | ActiveDC · CIFAR10 (0.5%), CIFAR100 (5%), ImageNet-1k (1%) | Top-1 accuracy: CIFAR10: 85.9, 87.3, 85.1, 81.0; CIFAR100: 64.5, 71.9, 65.3, 57.1; ImageNet: 54.7, 56.3, 56.1, 55.0 | Mixed |
| Increase ξ from -0.1 to 0.2 (but not to 0.3) (ablation) | ActiveDC · CIFAR10 (0.5%), CIFAR100 (5%), ImageNet-1k (1%) | Top-1 accuracy: CIFAR10: 85.5, 86.4, 86.9, 87.3, 70.2; CIFAR100: 65.3, 68.6, 71.5, 71.9, 61.8; ImageNet: 48.7, 50.6, 56.0, 56.3, 50.2 | Mixed |
| Add Tukey’s power transformation, then statistics calibration, then feature filtering (ablation) | ActiveDC · CIFAR10 (0.1%, 0.2%) | Top-1 accuracy: 47.1 -> 50.3 -> 59.5 -> 61.3; 64.5 -> 66.8 -> 70.6 -> 73.1 | Helped |
Surprising findings
- Generating more pseudo-labeled features (3x or 4x labeled count) decreases accuracy, attributed to uncorrected mislabeling.
- Traditional active learning methods (CoreSet, VAAL, LearnLoss, etc.) often underperform random selection in the active finetuning setting.
Limitations
- Performance depends on hyperparameters (alpha, xi, lambda, generated-feature count) tuned per dataset.
- Excess generated pseudo-labeled data degrades accuracy due to residual mislabeling.
Experimental setup
Datasets: CIFAR10 CIFAR100 ImageNet-1k
Benchmarks: CIFAR10 CIFAR100 ImageNet-1k
Metrics: Top-1 accuracy data selection time
Baselines: Random FDS (K-Center-Greedy) K-Means CoreSet VAAL LearnLoss TA-VAAL ALFA-Mix ActiveFT
Builds on: ActiveFT DINO DeiT Tukey's Ladder of Powers
Closest papers by meaning
- Boundary Matters: A Bi-Level Active Finetuning Method (NeurIPS 2024 · similarity 0.88)
- Composite Active Learning: Towards Multi-Domain Active Learning with Theoretical Guarantees (AAAI 2024 · similarity 0.84)
- A Cross-Domain Benchmark for Active Learning (NeurIPS 2024 · similarity 0.84)
- Uncertainty Meets Diversity: A Comprehensive Active Learning Framework for Indoor 3D Object Detection (CVPR 2025 · similarity 0.84)
- Plug and Play Active Learning for Object Detection (CVPR 2024 · similarity 0.83)
- Sketchy Moment Matching: Toward Fast and Provable Data Selection for Finetuning (NeurIPS 2024 · similarity 0.83)
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.