Research map / Optimization / Implicit bias and generalization
Data Cleansing for Models Trained with SGD
Satoshi Hara, Atsushi Nitanda, Takanori Maehara
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 SGD-step retracing to estimate influence instead of K&L | Influence estimation for LogReg and DNN · Adult, 20News, and MNIST | Kendall's tau: .93 (.02) vs .85 (.07); .94 (.05) vs .82 (.15); .95 (.02) vs .70 (.15) for LogReg; .75 (.10) vs .54 (.12); .45 (.12) vs | Improved |
| Use SGD-step retracing to estimate influence instead of K&L | Influence estimation for LogReg and DNN · Adult, 20News, and MNIST | Jaccard index: .80 (.10) vs .60 (.17); .79 (.15) vs .52 (.19); .83 (.10) vs .41 (.16) for LogReg; .59 (.16) vs .32 (.11); .25 (.08) vs | Improved |
| Remove the most influential instances identified by the proposed method | CNN trained on MNIST and CIFAR10 · MNIST and CIFAR10 test sets | misclassification rate: 25 cases in MNIST and 26 cases in CIFAR10 out of 30 experiments showed decreased misclassification rates | Improved |
| Retrace only one SGD epoch instead of all SGD steps (ablation) | Proposed influence estimation for data cleansing · MNIST and CIFAR10 | misclassification rate | No clear effect |
| Remove instances ranked by Autoencoder outlierness | CNN trained on MNIST and CIFAR10 · MNIST and CIFAR10 test sets | misclassification rate | No clear effect |
| Remove influential instances identified by the proposed method rather than by K&L | CNN data cleansing · MNIST and CIFAR10 | misclassification rate: 10 and 1000; 10 and 100 | Improved |
Closest papers by meaning
- Algorithms that Approximate Data Removal: New Results and Limitations (NeurIPS 2022 · similarity 0.85)
- Mixed-Sample SGD: an End-to-end Analysis of Supervised Transfer Learning (NeurIPS 2025 · similarity 0.82)
- On the Convergence of Loss and Uncertainty-based Active Learning Algorithms (NeurIPS 2024 · similarity 0.81)
- Hessian-Free Online Certified Unlearning (ICLR 2025 · similarity 0.81)
- Data Unlearning in Diffusion Models (ICLR 2025 · similarity 0.81)
- Most Influential Subset Selection: Challenges, Promises, and Beyond (NeurIPS 2024 · similarity 0.81)
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.
