Research map / Learning theory and algorithms / Kernels and regression theory
Scale-invariant Bayesian Neural Networks with Connectivity Tangent Kernel
SungYub Kim, Sihwan Park, Kyung-Su Kim, Eunho Yang
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 connectivity-based perturbations to measure sharpness | VGG-13 variants on CIFAR-10 · CIFAR-10 generalization-gap correlation | Kendall's τ: 0.755 -> 0.837 | Improved |
| Use connectivity-based perturbations to measure sharpness | VGG-13 variants on CIFAR-10 · CIFAR-10 generalization-gap correlation | Granulated Kendall's Ψ (average): 0.783 -> 0.900 | Improved |
| Use connectivity-based perturbations to measure sharpness | VGG-13 variants on CIFAR-10 · CIFAR-10 generalization-gap correlation | Conditional MI (K): 0.483 -> 0.539 | Improved |
| Replace Linearized Laplace with Connectivity Laplace | ResNet-18 · CIFAR-100 | NLL: 1.1673 ± 0.0093 -> 1.1307 ± 0.0042 | Improved |
| Replace Linearized Laplace with Connectivity Laplace | ResNet-18 · CIFAR-100 | ECE: 0.0632 ± 0.0010 -> 0.0524 ± 0.0009 | Improved |
| Replace Linearized Laplace with Connectivity Laplace | ResNet-18 · CIFAR-100 | Brier score: 0.3597 ± 0.0020 -> 0.3319 ± 0.0005 | Improved |
| Replace Linearized Laplace with Connectivity Laplace | ResNet-18 · CIFAR-100 OOD detection (SVHN) | AUC: 0.8066 ± 0.0120 -> 0.8423 ± 0.0204 | Improved |
| Use Connectivity Laplace instead of Linearized Laplace | ResNet-18 · CIFAR-100 | NLL: 1.0110 -> 1.1307 ± 0.0042 | Hurt |
| Use Connectivity Laplace instead of Linearized Laplace | ResNet-18 · CIFAR-100 | ECE: 0.0507 -> 0.0524 ± 0.0009 | Hurt |
| Use Connectivity Laplace instead of Linearized Laplace | ResNet-18 · CIFAR-100 | Brier score: 0.2740 -> 0.3319 ± 0.0005 | Hurt |
| Scale function-preserving parameters preceding batch normalization (ablation) | ResNet-18 with PAC-Bayes-CTK · CIFAR-10 and CIFAR-100 | PAC-Bayes-CTK bound: CIFAR-10: 0.0918 ± 0.0013 -> 0.0907 ± 0.0009; CIFAR-100: 0.2874 ± 0.0034 -> 0.2862 ± 0.0032 | No clear effect |
| Use connectivity-based damping rather than parameter-based damping | Laplace approximation with weight decay on batch-normalized networks · Function-preserving weight-decay scaling | Predictive uncertainty: LL uncertainty is amplified by 1/γ²; CL uncertainty is preserved | Improved |
Closest papers by meaning
- Normalized Flat Minima: Exploring Scale Invariant Definition of Flat Minima for Neural Networks Using PAC-Bayesian Analysis (ICML 2020 · similarity 0.92)
- Singular Bayesian Neural Networks (ICML 2026 · similarity 0.89)
- Flat Seeking Bayesian Neural Networks (NeurIPS 2023 · similarity 0.88)
- Generalization Bounds via Meta-Learned Model Representations: PAC-Bayes and Sample Compression Hypernetworks (ICML 2025 · similarity 0.86)
- Stable Minima of ReLU Neural Networks Suffer from the Curse of Dimensionality: The Neural Shattering Phenomenon (NeurIPS 2025 · similarity 0.86)
- A Symmetry-Aware Exploration of Bayesian Neural Network Posteriors (ICLR 2024 · similarity 0.86)
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.
