Research map / Learning theory and algorithms / Prediction and decision losses
Understanding Data Influence With Differential Approximation
Haoru Tan, Sitong Wu, Xiuzhe Wu, Wang Wang, Bo Zhao, Zeke Xie, Gui-Song Xia, Xiaojuan Qi
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 self-influence instead of validation influence (ablation) | Diff-In · Data cleaning CIFAR-100 (20%) | precision: 84.3 -> 91.4 | Improved |
| use self-influence instead of validation influence (ablation) | Diff-In · Data cleaning ImageNet-1K (20%) | precision: 47.2 -> 53.6 | Improved |
| replace TracIN with Diff-In (self-influence) | TracIN · Data cleaning Tiny-ImageNet (20%) | precision: 76.3 -> 88.2 | Improved |
| replace TracIN with Diff-In (self-influence) | TracIN · Data cleaning ImageNet-1K (20%) | precision: 46.4 -> 53.6 | Improved |
| replace SSSE with Diff-In | SSSE · Data deletion GSM8K | Accuracy@1: 85.7 -> 89.9 | Improved |
| replace TracIn with Diff-In | TracIn · Coreset selection Tiny-ImageNet (20%) | Accuracy@1: 24.2 -> 26.6 | Improved |
| replace TracIn with Diff-In | TracIn · Coreset selection ImageNet-1K (30%) | Accuracy@1: 64.4 -> 68.3 | Improved |
| replace Moderate with Diff-In | Moderate · Coreset selection CIFAR-100 (70%) | Accuracy@1: 71.8 -> 71.5 | Hurt |
| use all uniformly selected time steps instead of only the final one (ablation) | Diff-In-F · CC12M zero-shot classification (60%) | Acc@1: 34.4 -> 36.2 | Improved |
| replace CLIP-score selection with Diff-In | CLIP-score · CC12M zero-shot classification (60%) | Acc@1: 34.1 -> 36.2 | Improved |
| increase number of time steps m from 3 to 5 (ablation) | Diff-In · CC12M zero-shot classification | Acc@1: 24.3 -> 25.8 | Improved |
| increase number of time steps m from 5 to 50 (ablation) | Diff-In · CC12M zero-shot classification | Acc@1: 25.8 -> 26.5 | No clear effect |
Closest papers by meaning
- Capturing the Temporal Dependence of Training Data Influence (ICLR 2025 · similarity 0.88)
- Influence Functions for Scalable Data Attribution in Diffusion Models (ICLR 2025 · similarity 0.87)
- Outlier Gradient Analysis: Efficiently Identifying Detrimental Training Samples for Deep Learning Models (ICML 2025 · similarity 0.86)
- Final-Model-Only Data Attribution with a Unifying View of Gradient-Based Methods (NeurIPS 2025 · similarity 0.86)
- The Mirrored Influence Hypothesis: Efficient Data Influence Estimation by Harnessing Forward Passes (CVPR 2024 · similarity 0.86)
- DataInf: Efficiently Estimating Data Influence in LoRA-tuned LLMs and Diffusion Models (ICLR 2024 · similarity 0.86)
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.
