Research map / Label-efficient and robust learning / Recognition and long-tail
Data-centric Prediction Explanation via Kernelized Stein Discrepancy
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HD-Explain, an example-based prediction explanation method that uses a model-dependent Kernelized Stein Discrepancy kernel (built from input gradients and predicted log-probabilities) to retrieve top training points for a test input.
Problem
Example-based prediction explanation for neural classifiers is either computationally expensive (influence-based methods needing inverse Hessians or parameter perturbation) or coarse-grained (e.g., RPS returning class-level prototypes), so a precise, efficient, data-centric alternative is needed.
Approach
Uses the Stein operator-augmented kernel from Kernelized Stein Discrepancy to define a pairwise, model-conditioned data correlation between a test point (paired with its predicted label) and training points, with a per-sample score function (input gradient and log-prediction) that can be cached, avoiding parameter-space influence computation.
Key results
- Under noise injection, existing baselines reach at most about 10% hit rate while HD-Explain (and its last-layer variant) exceeds 80%.
- Baseline explainers achieve only 10%-50% coverage, while HD-Explain shows substantially higher coverage across test points.
- RPS had the lowest compute time; HD-Explain* ran faster than TracIn* and Influence Function, while full-model HD-Explain beat Influence Function on all datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HD-Explain or HD-Explain* instead of existing explanation methods | Example-based prediction explanation · Image classification datasets, Noise Injection | Hit Rate: >80% vs. ≤ 10% | Helped |
| Use HD-Explain instead of existing explanation methods | Example-based prediction explanation · Image classification datasets | Coverage: 10% - 50% | Helped |
| Use raw features rather than learned data representations (ablation) | HD-Explain · Image classification datasets, Horizontal Flip | Hit Rate | Hurt |
| Use IMQ kernel rather than RBF kernel (ablation) | HD-Explain · Image classification datasets, especially Horizontal Flip | Hit Rate | Helped |
| Use Linear kernel rather than RBF or IMQ kernel (ablation) | HD-Explain · Image classification datasets | Hit Rate | Hurt |
| Use Linear kernel rather than RBF or IMQ kernel (ablation) | HD-Explain · Image classification datasets | Execution time | Helped |
| Use HD-Explain* rather than full-model HD-Explain (ablation) | HD-Explain · CIFAR-10, mislabeled-sample retrieval | Precision@K and Recall@K | Helped |
Surprising findings
- Whole-model HD-Explain performs poorly on the data-debugging task, while its last-layer variant performs well.
- Horizontal flipping destroys raw-pixel similarity and degrades the full-model HD-Explain, while the last-layer variant is less affected.
Limitations
- Explanation quality depends on the model's prediction confidence and generalization; low-confidence predictions yield explanations from other classes.
- Kernel choice (Linear, RBF, IMQ) significantly affects performance, requiring empirical selection.
- Healthcare dataset explanations are not validated by domain experts; relaxations on input distribution are only argued heuristically for this context.
Experimental setup
Datasets: CIFAR-10 SVHN Brain Tumor MRI Ovarian Cancer histopathology Two Moons Rectangulars
Benchmarks: CIFAR-10 SVHN Brain Tumor MRI Ovarian Cancer histopathology
Metrics: Hit Rate Coverage Run Time Precision@K Recall@K (data debugging)
Baselines: Influence Function Representer Point Selection TracIn RBF kernel similarity
Builds on: Influence Function Representer Point Selection TracIn Kernelized Stein Discrepancy
Closest papers by meaning
- A Learning Theoretic Perspective on Local Explainability (ICLR 2021 · similarity 0.87)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.85)
- Efficient and Accurate Explanation Estimation with Distribution Compression (ICLR 2025 · similarity 0.85)
- Faithful and Efficient Explanations for Neural Networks via Neural Tangent Kernel Surrogate Models (ICLR 2024 · similarity 0.85)
- ReX: A Framework for Incorporating Temporal Information in Model-Agnostic Local Explanation Techniques (AAAI 2025 · similarity 0.84)
- Selective Explanations (NeurIPS 2024 · similarity 0.84)
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.