Research map / Architectures and efficiency / RNNs and normalization
SmoothHess: ReLU Network Feature Interactions via Stein's Lemma
Max Torop, Aria Masoomi, Davin Hill, Kivanc Kose, Stratis Ioannidis, Jennifer Dy
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 SmoothHess + SmoothGrad instead of SoftPlus Hessian + Gradient | Taylor-expansion explainer · MNIST class logit, eps=0.25 | P_MSE: 1.2e-6 -> 9.6e-7 | Improved |
| Use SmoothHess + SmoothGrad instead of SoftPlus Hessian + Gradient | Taylor-expansion explainer · CIFAR10 class logit, eps=0.25 | P_MSE: 2.1e-3 -> 9.8e-4 | Improved |
| Use SmoothHess + SmoothGrad instead of SoftPlus Hessian + Gradient | Taylor-expansion explainer · CIFAR10 penultimate neuron, eps=0.25 | P_MSE: 1.1e-5 -> 8.1e-7 | Improved |
| Add the Hessian term to SmoothGrad (first order to second order) (ablation) | Gaussian smoothing · CIFAR10 class logit, eps=0.25 | P_MSE: 1.3e-2 -> 9.8e-4 | Improved |
| Add the Hessian term to SoftPlus gradient (ablation) | SoftPlus smoothing · MNIST class logit, eps=1.00 | P_MSE: 3.9e-4 -> 8.1e-5 | Improved |
| Smooth the gradient with SoftPlus vs SmoothGrad (first-order only) | first-order explainers · MNIST class logit, eps=0.25 | P_MSE: 4.6e-6 vs 4.5e-6 | No clear effect |
| Use SmoothGrad instead of vanilla gradient | first-order explainer · MNIST class logit, eps=0.25 | P_MSE: 4.2e-3 -> 4.5e-6 | Improved |
| Choose sigma from only 3 values (eps/sqrt(d) rule) vs SoftPlus beta from over 100 values | SmoothHess · P_MSE, all datasets | P_MSE | Improved |
| Use SmoothHess + SmoothGrad attack instead of SoftPlus Hessian + Gradient attack | second-order Taylor attack · MNIST eps=0.75 | post-hoc accuracy: 51.5 -> 48.0 | Improved |
| Use SmoothHess + SmoothGrad attack instead of SoftPlus Hessian + Gradient attack | second-order Taylor attack · CIFAR10 eps=0.4 | post-hoc accuracy: 23.5 -> 15.0 | Improved |
| Use SmoothHess + SmoothGrad attack instead of vanilla gradient attack | Taylor attack · CIFAR10 eps=0.4 | post-hoc accuracy: 34.0 -> 15.0 | Improved |
| Add Hessian term to SmoothGrad for attacks at large radius (ablation) | SmoothGrad attack · CIFAR10 eps=1.0 | post-hoc accuracy: 0.0 -> 4.5 | Hurt |
Closest papers by meaning
- Explaining Explanations: Axiomatic Feature Interactions for Deep Networks (JMLR 2021 · similarity 0.85)
- The inductive bias of ReLU networks on orthogonally separable data (ICLR 2021 · similarity 0.85)
- On the Complexity-Faithfulness Trade-off of Gradient-Based Explanations (ICCV 2025 · similarity 0.85)
- Make Haste Slowly: A Theory of Emergent Structured Mixed Selectivity in Feature Learning ReLU Networks (ICLR 2025 · similarity 0.84)
- Simplicity Bias and Optimization Threshold in Two-Layer ReLU Networks (ICML 2025 · similarity 0.83)
- Expressivity of ReLU-Networks under Convex Relaxations (ICLR 2024 · similarity 0.83)
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.
