Research map / Learning theory and algorithms / Prediction and decision losses
DeepCSHAP: Utilizing Shapley Values to Explain Deep Complex-Valued Neural Networks
Florian Eilers, Xiaoyi Jiang
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 |
|---|---|---|---|
| use sum of real and imaginary parts instead of absolute value for complex gradient saliency (ablation) | gradient-based explanation methods · MNIST | change in log-odds after masking 20% pixels | Improved |
| use CReLU-projection variant instead of zReLU variant in Guided Backpropagation (ablation) | Guided Backpropagation · MNIST | change in log-odds after masking 20% pixels | Improved |
| use DeepCSHAP instead of complex gradient-based methods | explanation of complex CNN · MNIST | change in log-odds after masking 20% pixels | Improved |
| use DeepCSHAP instead of complex gradient-based methods | explanation of complex ResNet18 · S1SLC_CVDL two-patch channel attribution | fraction of attribution in correct channels | Improved |
| use absolute value of gradient instead of R+I (ablation) | gradient-based methods · S1SLC_CVDL two-patch channel attribution | fraction of attribution in correct channels: median scores around 0.5 vs around 1 with R+I | Hurt |
| apply Guided Backpropagation variants | gradient-based methods · S1SLC_CVDL two-patch channel attribution | fraction of attribution in correct channels | Hurt |
| replace multiplier denominators below 1e-6 with Wirtinger derivative for numerical stability | DeepCSHAP · MNIST / S1SLC_CVDL | Local Accuracy Error: 9.8 x 10^-6 on MNIST, 3.5 x 10^-3 on S1SLC_CVDL | Mixed |
| use integrated gradients with 5 integrands instead of plain gradient | Grad · MNIST | runtime (ms): 5.0 -> 22.6 | Hurt |
| use DeepCSHAP instead of gradient methods | explanation cost · S1SLC_CVDL | runtime (ms): 57.2 -> 380 | Hurt |
Closest papers by meaning
- SHAP values via sparse Fourier representation (NeurIPS 2025 · similarity 0.84)
- OperatorSHAP: Fast and Accurate Shapley Value Estimation for Neural Operators (NeurIPS 2026 · similarity 0.83)
- SHAP Meets Tensor Networks: Provably Tractable Explanations with Parallelism (NeurIPS 2025 · similarity 0.83)
- Exact Computation of Any-Order Shapley Interactions for Graph Neural Networks (ICLR 2025 · similarity 0.82)
- Towards Trustable SHAP Scores (AAAI 2025 · similarity 0.82)
- Prediction via Shapley Value Regression (ICML 2025 · similarity 0.82)
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.
