Research map / Label-efficient and robust learning / Self-supervised and contrastive
A Mechanism-Driven Theory of Phase Transitions in Active Learning
Julia Machnio, Mads Nielsen, MOSTAFA MEHDIPOUR GHAZI
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a self-supervised SimCLR backbone instead of supervised training | segmented regression phase detection · CIFAR-100 | second transition tau2 (labels): 37,900 -> 11,000 | Improved |
| increase number of segments from K=2 to K=3 (ablation) | segmented regression · CIFAR-10 | BIC: -42188 -> -43235 | Improved |
| increase number of segments from K=3 to K=5 (ablation) | segmented regression · CIFAR-10 | SSE: 0.09 -> 0.06 | Improved |
| increase segments from K=1 to K=3 (ablation) | segmented regression · ISIC | R^2: 0.828 -> 0.937 | Improved |
| increase segments from K=3 to K=4 (ablation) | segmented regression · CIFAR-10 | R^2: 0.998 -> 0.998 | No clear effect |
| use TypiClust (representativeness) in the first phase | active learning strategies · CIFAR-100 (phase I, <=5k labels) | accuracy: ~2% accuracy advantage | Improved |
| switch among TypiClust, Coreset, Uncertainty at detected transitions (TCU) | static acquisition baselines · CIFAR benchmarks | accuracy | Improved |
| use BYOL embeddings instead of MoCov3 for TypiClust | TypiClust · ImageNet-100 low budget | accuracy | Hurt |
| use representativeness or coverage-based selection instead of uncertainty at scale | TypiClust · CIFAR-10 runtime | runtime ratio: 6.79 | Hurt |
| use pretrained SSL representations instead of training from scratch | active learning pipeline · training cost | training speedup: 30x to over 160x | Improved |
Closest papers by meaning
- Transductive Active Learning: Theory and Applications (NeurIPS 2024 · similarity 0.85)
- What needs to go right for an induction head? A mechanistic study of in-context learning circuits and their formation (ICML 2024 · similarity 0.85)
- Why Diffusion Models Don’t Memorize: The Role of Implicit Dynamical Regularization in Training (NeurIPS 2025 · similarity 0.85)
- High-Dimensional Learning Dynamics of Attention-Indexed Models (NeurIPS 2026 · similarity 0.85)
- A Closer Look at Model Collapse: From a Generalization-to-Memorization Perspective (NeurIPS 2025 · similarity 0.84)
- Understanding and Leveraging the Learning Phases of Neural Networks (AAAI 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.
