Research map / Learning theory and algorithms / Prediction and decision losses
Influence Functions in Deep Learning Are Fragile
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 |
|---|---|---|---|
| add weight decay (ablation) | Iris feed-forward network · Iris | Spearman correlation: 0.508 -> 0.97 | Helped |
| increase network depth (ablation) | Iris feed-forward network · Iris | Spearman correlation | Hurt |
| increase network width from 8 to 50 (ablation) | Iris feed-forward network · Iris | Spearman correlation: 0.82 -> 0.56 | Hurt |
| use stochastic estimation instead of the exact inverse-Hessian vector product (ablation) | Influence functions · Iris | Spearman correlation | Hurt |
| add weight decay (ablation) | Small CNN influence estimates · MNIST, highest-loss test point | Spearman correlation: 0.35 -> 0.87 | Helped |
| evaluate lower-ranked training points around the 30th percentile instead of the top influential points (ablation) | Small CNN influence estimates · small MNIST | correlation with ground-truth influence | Hurt |
| increase weight decay beyond 0.01 (ablation) | Small CNN influence estimates · small MNIST | correlation with ground-truth influence | Hurt |
| use a test point at the 50th percentile of test losses rather than the highest-loss test point (ablation) | Influence functions · MNIST and CIFAR-10 | correlation with ground-truth influence | Mixed |
| remove weight decay (ablation) | LeNet influence estimates · CIFAR-10, highest-loss test point | Spearman correlation: 0.69 -> 0.09 | Hurt |
| increase weight decay from 0.03 to 0.075 (ablation) | Deeper Iris network · Iris | Spearman correlation: 0.47 -> 0.72 | Helped |
| use single-example influence estimates at ImageNet scale | ResNet-50 influence estimates · ImageNet | Pearson and Spearman correlation: less than $0.15$ | Hurt |
| use a larger group of training samples rather than a single sample | Influence estimates · ImageNet | reliability of ground-truth estimates | Mixed |
Closest papers by meaning
- Towards Robust Influence Functions with Flat Validation Minima (ICML 2025 · similarity 0.91)
- Influence Functions for Edge Edits in Non-Convex Graph Neural Networks (NeurIPS 2025 · similarity 0.89)
- A Versatile Influence Function for Data Attribution with Non-Decomposable Loss (ICML 2025 · similarity 0.88)
- Outlier Gradient Analysis: Efficiently Identifying Detrimental Training Samples for Deep Learning Models (ICML 2025 · similarity 0.87)
- Influence Functions for Scalable Data Attribution in Diffusion Models (ICLR 2025 · 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.