Research map / Learning theory and algorithms / Prediction and decision losses
Understanding Black-box Predictions via Influence Functions
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 |
|---|---|---|---|
| replace exact hinge with SmoothHinge t=0.001 for influence (ablation) | influence functions on linear SVM · 1s vs 7s MNIST | Pearson's R: 0.95 | Helped |
| increase smoothing temperature from 0.001 to 0.1 (ablation) | SmoothHinge influence · 1s vs 7s MNIST | Pearson's R: 0.95 -> 0.91 | Hurt |
| apply influence to non-converged non-convex CNN with damping 0.01 | influence functions · MNIST CNN (500k iters) | Pearson's R: 0.86 | No clear effect |
| use stochastic approximation of inverse HVP (r=10, t=5,000) instead of exact CG (ablation) | influence functions · 10-class MNIST logistic regression | agreement with leave-one-out | No clear effect |
| reduce stochastic repeats to r=1 (ablation) | stochastic inverse HVP · 10-class MNIST | identification of influential points | Mixed |
| perturb 1 training image with influence-based attack | Inception top layer · dogs vs fish | test predictions flipped: 335 of 591 (57%) | Helped |
| perturb 2 training images instead of 1 (ablation) | influence training attack · dogs vs fish | test predictions flipped: 57% -> 77% | Helped |
| perturb 10 training images (ablation) | influence training attack · dogs vs fish | test predictions flipped: 77% -> all but 1 of 591 | Helped |
| inspect parameter coefficients instead of influence functions | debugging baseline · hospital readmission child domain mismatch | identifying responsible features: 14/127 features had a larger coefficient | Hurt |
| use influence function ranking to identify responsible training points | influence functions · hospital readmission | identification of 4 child training points: 30-40 times as influential | Helped |
| prioritize flipped-label checking by self-influence rather than train loss or random | influence functions · Enron1 spam | test accuracy and flipped data detected | Helped |
| train a further 500k iterations with full passes | non-converged CNN · MNIST subset | train loss: 0.14 -> 0.12 | Helped |
Closest papers by meaning
- Towards Robust Influence Functions with Flat Validation Minima (ICML 2025 · similarity 0.88)
- 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 0.87)
- Most Influential Subset Selection: Challenges, Promises, and Beyond (NeurIPS 2024 · similarity 0.86)
- Rescaled Influence Functions: Accurate Data Attribution in High Dimension (NeurIPS 2025 · 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.