Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
A Bayesian Approach To Analysing Training Data Attribution In Deep Learning
Elisa Nguyen, Minjoon Seo, Seong Joon Oh
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 batch composition (DE-Batch) instead of initialisation (DE-Init) as the randomness source (ablation) | LOO ground truth · MNIST3 CNN2-L | mean p-value of TDA: 0.331 -> 0.025 | Improved |
| use DE-Batch instead of DE-Init as randomness source (ablation) | LOO ground truth · CIFAR10 CNN2-L | mean p-value of TDA: 0.692 -> 0.487 | Improved |
| use DE-Batch instead of DE-Init as randomness source (ablation) | Influence functions · MNIST3 CNN2-L | mean p-value of TDA: 0.352 -> 0.000 | Improved |
| increase model complexity from CNN2-L to ViT+LoRA (ablation) | LOO ground truth · MNIST3 | mean p-value of TDA: 0.025 -> 0.786 | Hurt |
| add a third layer (CNN2-L to CNN3-L) (ablation) | LOO ground truth · MNIST3 | mean p-value of TDA: 0.025 -> 0.370 | Hurt |
| increase model complexity from CNN2-L to ViT+LoRA (ablation) | LOO ground truth · CIFAR10 | mean p-value of TDA: 0.623 -> 0.777 | Hurt |
| use Grad-Cos instead of influence functions | TDA estimator · MNIST3 SWA+DE-Init | mean p-value of TDA: 0.352 -> 0.003 | Mixed |
| train CNN2-L on full MNIST (60,000) rather than small subset (ablation) | LOO ground truth · MNIST | mean p-value of TDA: 0.331 -> 0.761 | Hurt |
| increase number of SWA/DE posterior samples (ablation) | LOO ground truth · MNIST3, CIFAR10 | LOO p-value stability | Improved |
| restrict correlation analysis to low-noise train-test pairs (LOO p<=0.05) (ablation) | TDA vs LOO correlation · MNIST3, CIFAR10 | Spearman correlation of mean with LOO | Improved |
| treat TDA as probabilistic (posterior mean) rather than deterministic point estimate | TDA self-influence mislabel detection · MNIST3, CIFAR10 | fraction of mislabels found | Improved |
| increase training set size (ablation) | TDA stability (CNN2-L) · MNIST3, CIFAR10 | mean p-value of TDA: MNIST3 SWA+DE-Init LOO 0.058 -> 0.714 -> 0.374 (30, 90, 180 samples) | Mixed |
Closest papers by meaning
- quanda: An Interpretability Toolkit for Training Data Attribution Evaluation (NeurIPS 2026 · similarity 0.90)
- Final-Model-Only Data Attribution with a Unifying View of Gradient-Based Methods (NeurIPS 2025 · similarity 0.89)
- Distributional Training Data Attribution: What do Influence Functions Sample? (NeurIPS 2025 · similarity 0.88)
- Better Training Data Attribution via Better Inverse Hessian-Vector Products (NeurIPS 2025 · similarity 0.87)
- Training Data Attribution via Approximate Unrolling (NeurIPS 2024 · similarity 0.86)
- LoRIF: Low-Rank Influence Functions for Scalable Training Data Attribution (NeurIPS 2026 · 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.
