Research map / Optimization / Optimizers
Estimating Training Data Influence by Tracing Gradient Descent
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 TracInCP rather than influence functions or representer points to rank suspected mislabels | CIFAR-10 ResNet-56 · CIFAR-10 mislabelled training examples, first 20% of ranking | fraction of mislabelled data recovered: more than 80% vs less than 50% | Helped |
| Fix mislabelled examples identified by TracIn rather than by the compared methods | CIFAR-10 ResNet-56 · CIFAR-10 with injected label errors | test accuracy improvement after fixing identified labels | Helped |
| Use TracIn rather than influence functions or representer points to rank suspected mislabels | MNIST 3-layer model · MNIST mislabelled training examples | fraction of mislabelled examples retrieved | Helped |
| Use step-level first-order TracIn instead of checkpoint-based TracInCP | MNIST 3-layer model · MNIST mislabelled training examples | speed of mislabelled-example recovery | Helped |
| Approximate per-step loss change with first-order TracIn (ablation) | MNIST 3-layer model · MNIST step-level influence evaluation | Pearson correlation with test-loss change: 0.978 | Helped |
| Use checkpoints selected at high loss-reduction points rather than evenly spaced checkpoints (ablation) | TracInCP · MNIST influence-score approximation | correlation with first-order TracIn scores | Helped |
| Add more checkpoints to TracInCP (ablation) | TracInCP · MNIST influence-score approximation | correlation with first-order TracIn scores | Helped |
| Weight checkpoints equally instead of weighting by learning rate (ablation) | TracInCP · CIFAR-10 mislabelled-example identification | fraction of mislabelled data recovered | No clear effect |
| Use only the final checkpoint rather than checkpoints across training (ablation) | TracInCP · Influence estimation | influence-score quality | Hurt |
Closest papers by meaning
- LoRIF: Low-Rank Influence Functions for Scalable Training Data Attribution (NeurIPS 2026 · similarity 0.85)
- Final-Model-Only Data Attribution with a Unifying View of Gradient-Based Methods (NeurIPS 2025 · similarity 0.84)
- The Mirrored Influence Hypothesis: Efficient Data Influence Estimation by Harnessing Forward Passes (CVPR 2024 · similarity 0.83)
- Outlier Gradient Analysis: Efficiently Identifying Detrimental Training Samples for Deep Learning Models (ICML 2025 · similarity 0.83)
- Enhancing Training Data Attribution with Representational Optimization (NeurIPS 2025 · similarity 0.83)
- Grokking as the transition from lazy to rich training dynamics (ICLR 2024 · similarity 0.83)
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.