Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Towards Accelerated Model Training via Bayesian Data Selection
Zhijie Deng, Peng Cui, Jun Zhu
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 Bayesian generalization-loss-based online data selection instead of RHO-LOSS | ResNet18 training on CIFAR-100 · CIFAR-100 | epochs to reach 40.0% accuracy: 48 -> 32 | Improved |
| use Bayesian generalization-loss-based online data selection instead of RHO-LOSS | ResNet18 training on WebVision-200 · WebVision-200, ILSVRC12 validation | final accuracy: 39% -> 56% | Improved |
| replace the zero-shot predictor with the RHO-LOSS validation model (ablation) | Proposed method · CIFAR-100 | epochs to reach 40.0% test accuracy: 32 -> 30 | Improved |
| replace the zero-shot predictor with the RHO-LOSS validation model (ablation) | Proposed method · CIFAR-100 | final accuracy: 63% -> 63% | No clear effect |
| use only the zero-shot predictor term (α=0) (ablation) | Proposed method · CIFAR-100 | final accuracy | Hurt |
| increase α enough to downweight the zero-shot predictor term (ablation) | Proposed method · CIFAR-100 | training speed | Hurt |
| vary the effective-data parameter n_e (ablation) | Proposed method · CIFAR-100 | training performance | No clear effect |
| use an appropriate CLIP softmax temperature τ (ablation) | Proposed method · CIFAR-100 | training performance | Improved |
| use CLIP RN50 rather than the higher-zero-shot-accuracy ViT-B/16 predictor (ablation) | Proposed method · CIFAR-100 | training speedup | No clear effect |
| replace the ResNet18 training backbone with a pretrained ViT-B/16 (ablation) | Proposed method · WebVision-200 | training performance | Improved |
Closest papers by meaning
- Diversified Batch Selection for Training Acceleration (ICML 2024 · similarity 0.87)
- Data Mixture Optimization: A Multi-fidelity Multi-scale Bayesian Framework (NeurIPS 2025 · similarity 0.87)
- DsDm: Model-Aware Dataset Selection with Datamodels (ICML 2024 · similarity 0.87)
- A Bayesian Approach to Data Point Selection (NeurIPS 2024 · similarity 0.85)
- Predictive Data Selection: The Data That Predicts Is the Data That Teaches (ICML 2025 · similarity 0.85)
- GREATS: Online Selection of High-Quality Data for LLM Training in Every Iteration (NeurIPS 2024 · similarity 0.85)
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.
