Research map / Architectures and efficiency / RNNs and normalization
Hidden Monotonicity: Explaining Deep Neural Networks via their DC Decomposition
Jakob Paul Zimmermann, Georg Loho
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 LayerCAM with SplitCAM on the split network | VGG16 (pretrained) · ImageNet-S-50 test | Pointing Game: 0.855 -> 0.938 | Improved |
| replace gamma-LRP with SplitLRP | VGG16 (pretrained) · ImageNet-S-50 test | Pointing Game: 0.680 -> 0.871 | Improved |
| replace LayerCAM with SplitCAM on the split network | ResNet18 (pretrained) · ImageNet-S-50 test | Pointing Game: 0.971 -> 0.972 | Improved |
| replace LayerCAM with SplitCAM on the split network | ResNet18 (pretrained) · ImageNet-S-50 test | Selectivity (lower is better): 3.690 -> 3.503 | Improved |
| use backward stabilization alpha=0.4 instead of alpha=0 for SplitGrad (ablation) | SplitGrad (26, g), VGG16 · ImageNet-S-50 test | Pointing Game: 0.401 -> 0.947 | Improved |
| use alpha=0.4 instead of alpha=0.5 (scaled original gradients) for SplitGrad (ablation) | SplitGrad (26, g), VGG16 · ImageNet-S-50 test | Pointing Game: 0.898 -> 0.947 | Improved |
| use alpha=0.4 instead of alpha=0.5 for SplitCAM (ablation) | SplitCAM (26, +, g), VGG16 · ImageNet-S-50 test | Pointing Game: 0.868 -> 0.938 | Improved |
| use alpha=0 (no stabilization) instead of intermediate alpha (ablation) | SplitGrad (26, g), VGG16 · ImageNet-S-50 test | Maximum Sensitivity (lower is better): 0.616 -> 0.028 | Improved |
| very small alpha (0.3) for SplitCAM (ablation) | SplitCAM (26, +, g), VGG16 · ImageNet-S-50 test | Pointing Game: 0.938 -> 0.613 | Hurt |
| directly train as difference of convex or monotone networks (DIC/DM) on MLP maxout | MNIST mlp maxout · MNIST test | accuracy | Improved |
| invert the input to the h-stream (DICIN/DMIN) (ablation) | DIC/DM models on MNIST · MNIST | explanation separation (qualitative) | Improved |
| use winner-takes-all max pooling instead of the convexity-preserving identity (ablation) | Split networks · ImageNet-S-50 | monotonicity and convexity of streams | Hurt |
Closest papers by meaning
- Enhancing Interpretability for Vision Models via Shapley Value Optimization (AAAI 2026 · similarity 0.86)
- Counterexample-Guided Learning of Monotonic Neural Networks (NeurIPS 2020 · similarity 0.86)
- Q-SENN: Quantized Self-Explaining Neural Networks (AAAI 2024 · similarity 0.85)
- On the explainable properties of 1-Lipschitz Neural Networks: An Optimal Transport Perspective (NeurIPS 2023 · similarity 0.85)
- Counterfactual-based Saliency Map: Towards Visual Contrastive Explanations for Neural Networks (ICCV 2023 · similarity 0.85)
- Explainable Saliency: Articulating Reasoning with Contextual Prioritization (CVPR 2025 · similarity 0.85)
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.
