Research map / Label-efficient and robust learning / Recognition and long-tail
SLICE: Stabilized LIME for Consistent Explanations for Image Classification
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 SLICE, a stabilized variant of LIME for image classifiers that removes unstable superpixels via sign-entropy-based feature elimination and perturbs images with an adaptively chosen Gaussian blur, improving explanation consistency (new CCM metric) and fidelity over LIME and BayLIME.
Problem
LIME explanations for image classifiers are inconsistent across runs for the same model and image, with superpixels flipping sign and changing importance rank.
Approach
Combine a sign-entropy-based iterative feature elimination (SEFE, using bootstrapped ridge regression and KDE) with an adaptive Gaussian-blur perturbation (Ada-Blur) whose sigma is chosen by maximizing adjusted R-squared of the surrogate, plus a new consistency metric CCM combining sign-flip entropy and rank-biased overlap.
Key results
- SLICE achieves significantly higher CCM consistency than LIME and BayLIME across 200 image-model pairs (Wilcoxon tests, p-values as low as 1e-9).
- SLICE attains a probability of negative AOPC (Pr-) of 0.0 on most dataset-model settings for insertion and deletion, versus 0.25 to 0.64 for LIME and BayLIME.
- SLICE is substantially slower than LIME: about 88.9 s vs 19.5 s per image for InceptionV3 and 44.7 s vs 11.7 s for ResNet50.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SLICE instead of LIME or BayLIME | Image classification explanations · Oxford-IIIT Pets and PASCAL VOC 2007 with InceptionV3 and ResNet50 | consistency (CCM) | Helped |
| Use SLICE instead of LIME or BayLIME | Image classification explanations · Oxford-IIIT Pets and PASCAL VOC 2007 with InceptionV3 and ResNet50 | AOPC fidelity: 0.001; 0.001; 0.002; 0.006; 0.002; 0.001; 0.003; 0.01; 0.01; 0.003; 0.01; 0.01; 0.005; 0.003; 0.006; 0.004 | Helped |
| Use SLICE instead of LIME or BayLIME | Image classification explanations · Oxford-IIIT Pets and PASCAL VOC 2007 with InceptionV3 and ResNet50 | insertion AUC: 0.008; 0.004; 0.002; 0.008; 0.019; 0.013; 0.012; 0.015 | Helped |
| Use SLICE instead of LIME or BayLIME | Image classification explanations · Oxford-IIIT Pets and PASCAL VOC 2007 with InceptionV3 and ResNet50 | deletion AUC: 0.005; 0.005; 0.007; 0.01; 0.006; 0.006; 0.007; 0.008 | Helped |
| Use SLICE instead of LIME or BayLIME | Image classification explanations · Oxford-IIIT Pets and PASCAL VOC 2007 with InceptionV3 and ResNet50 | runtime: 19.48s; 19.70s; 88.94; 11.73; 11.87; 44.73 | Hurt |
| Add Ada-Blur to SLICE FE (ablation) | SLICE · Oxford-IIIT Pets and PASCAL VOC 2007 with InceptionV3 and ResNet50 | consistency (CCM) | Helped |
| Add SEFE to SLICE blur (ablation) | SLICE · Oxford-IIIT Pets and PASCAL VOC 2007 with InceptionV3 and ResNet50 | consistency (CCM) | Helped |
Surprising findings
- Only one of 200 image-model combinations had no superpixels selected as stable by SEFE.
- Ada-Blur alone (without feature elimination) gives higher CCM than LIME but slightly lower than full SLICE, while SEFE alone with zeroing perturbation is weaker.
Limitations
- SLICE is roughly 4-5 times slower than LIME per image.
- Reliability of fidelity metrics (AOPC, insertion/deletion) was not investigated.
- Evaluated only on two pretrained ImageNet classifiers and 50 images per dataset.
Experimental setup
Datasets: Oxford-IIIT Pet PASCAL VOC 2007
Benchmarks: Oxford-IIIT Pet PASCAL VOC 2007
Metrics: CCM (combined consistency metric) ASFE ARS (rank biased overlap) AOPC insertion AUC deletion AUC Pr- runtime
Baselines: LIME BayLIME
Builds on: LIME BayLIME Rank Biased Overlap
Closest papers by meaning
- Using Stratified Sampling to Improve LIME Image Explanations (AAAI 2024 · similarity 0.90)
- What does LIME really see in images? (ICML 2021 · similarity 0.89)
- Locally Invariant Explanations: Towards Stable and Unidirectional Explanations through Local Invariant Learning (NeurIPS 2023 · similarity 0.88)
- I-CEE: Tailoring Explanations of Image Classification Models to User Expertise (AAAI 2024 · similarity 0.85)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.84)
- Probabilistic Conceptual Explainers: Trustworthy Conceptual Explanations for Vision Foundation Models (ICML 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.