Research map / Label-efficient and robust learning / Recognition and long-tail
Evaluating Sample Utility for Efficient Data Selection by Mimicking Model Weights
Tzu-Heng Huang, Manjot Bilkhu, John Cooper, Frederic Sala, Javier Movellan
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 |
|---|---|---|---|
| re-weight batch samples by mimic score | mini-batch SGD · six noisy image datasets, noise 0.5 (avg) | test accuracy (%): 69.46 -> 74.53 | Improved |
| weight samples by gradient norm (GraNd) | mini-batch SGD · six noisy image datasets, noise 0.5 (avg) | test accuracy (%): 69.46 -> 53.36 | Hurt |
| use mimic score instead of Rho-Loss excess loss | Rho-Loss · six noisy image datasets, noise 0.6 (avg) | test accuracy (%): 71.93 -> 73.01 | Improved |
| use mimic score with hold-out reference model | Rho-Loss · CIFAR100 50% noise | test accuracy (%): 75.11 -> 75.81 | Improved |
| use pretrained reference weights instead of a labeled validation set (ablation) | Grad-Mimic · six datasets, 50% noise | test accuracy | Improved |
| add mimic-score re-weighting to CLIP training | CLIP from scratch · DataComp small | 38-task average: 0.131 -> 0.146 | Improved |
| add mimic-score re-weighting to CLIP training | CLIP from scratch · DataComp medium | 38-task average: 0.254 -> 0.258 | Improved |
| use image-encoder last layer instead of text-encoder as reference (ablation) | Grad-Mimic · DataComp small | 38-task average: 0.139 -> 0.145 | Improved |
| mimic-score training convergence | CLIP from scratch · DataComp | training steps: 20.7% fewer steps | Improved |
| remove mimic-score bottom 5% from CLIP Score top 30% filter | CLIP Score filter · DataComp medium | 38-task average: 0.321 -> 0.323 | Improved |
| remove mimic-score bottom 20% from CLIP Score filter | CLIP Score filter · DataComp small | 38-task average: 0.165 -> 0.164 | No clear effect |
| use mimic score instead of GraNd/AGRA/influence for cost | Grad-Mimic · per-batch overhead | compute: 2.6x lower overhead | Improved |
Closest papers by meaning
- Scaling Laws for Data Filtering-- Data Curation cannot be Compute Agnostic (CVPR 2024 · similarity 0.86)
- A CLIP-Powered Framework for Robust and Generalizable Data Selection (ICLR 2025 · similarity 0.85)
- Filter Like You Test: Data-Driven Data Filtering for CLIP Pretraining (NeurIPS 2025 · similarity 0.85)
- When Dynamic Data Selection Meets Data Augmentation: Achieving Enhanced Training Acceleration (ICML 2025 · similarity 0.85)
- DELT: A Simple Diversity-driven EarlyLate Training for Dataset Distillation (CVPR 2025 · similarity 0.84)
- A Bitter Lesson for Data Filtering (NeurIPS 2026 · 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.
