SliceEvidence
Does the explanation survive a different test? 先检查证据，再判断原因。
200 real MNIST test images · frozen CNN · 6,400 precomputed occlusion runs · no model training
2 · Inspect the same image
Rose intensity = positive probability drop
3 · Change the evidence setting
Choose a 7 × 7 patch. The target remains the original predicted class.
Paired response / 配对结果
Positive Δp means masking lowered the original class probability; negative means it increased. Values are uncalibrated model outputs.
4 · Record a bounded conclusion
A changed prediction does not establish a real-world cause. Gray patches can be out of distribution; either fill can erase a defining stroke.
Attributed schematic baseline: summary + instance heatmap, inspired by AttributionScanner. This ablation hides paired evidence; it does not reproduce or benchmark the full original system.
Method, provenance, and limitations / 方法与边界
Data: ylecun/mnist, test rows 0–199. Records are real handwriting. Masked images are programmatically derived stress tests, not independent observations.
Model: onnxmodelzoo/mnist-12. Occlusion uses a 4 × 4 grid of non-overlapping 7 × 7 patches, fills 0 and 0.5, and a fixed original predicted class. Probability differences are computed by actual frozen CNN inference. There is no GradCAM, feature inversion, label propagation, retraining, or causal inference in this prototype.
Minimum support is a user-set review heuristic, not a statistical guarantee. Slice membership is selected on black-fill responses; agreement between two fills is not independent validation. No accuracy improvement or learner benefit has been established.
Adaptation: Xuan et al., AttributionScanner, IEEE VIS 2025 / TVCG 31(10), 7436–7447. Publication. Related baseline sensitivity: Sturmfels et al. (2020).
Concept map / 改进逻辑
Research prototype v0.1 · AI-assisted draft for Yizhou Zhang · human review and course deployment pending. Evidence decisions stay in memory until exported; no private input is sent anywhere.