Research map / Learning theory and algorithms / Kernels and regression theory
Non-vacuous Generalization Bounds at the ImageNet Scale: a PAC-Bayesian Compression Approach
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 compression length and weight-perturbation robustness in the PAC-Bayes bound | LeNet-5 · MNIST | test error bound: < 46% | Helped |
| Use compression length and weight-perturbation robustness in the PAC-Bayes bound | MobileNet 0.5 · ImageNet | top-1 test error bound: < 96.5% | Helped |
| Use compression length and weight-perturbation robustness in the PAC-Bayes bound | MobileNet 0.5 · ImageNet | top-5 test error bound: < 89% | Helped |
| Apply the simple Occam bound to the compressed SqueezeNet model | SqueezeNet · ImageNet | test error bound: 98.6% | Helped |
| Add Gaussian noise to the non-zero weights (ablation) | MobileNet 0.5 · ImageNet training | top-1 accuracy: 67% -> 65% | Hurt |
| Prune ResNet-56 to 50% sparsity (ablation) | ResNet-56 · CIFAR-10 with randomized labels | training accuracy | No clear effect |
| Increase label randomization while compressing beyond 50% sparsity (ablation) | ResNet-56 · CIFAR-10 with randomized labels | training accuracy | Mixed |
Closest papers by meaning
- Stronger Generalization Bounds for Deep Nets via a Compression Approach (ICML 2018 · similarity 0.92)
- PAC-Bayes Compression Bounds So Tight That They Can Explain Generalization (NeurIPS 2022 · similarity 0.91)
- Generalization Bounds via Meta-Learned Model Representations: PAC-Bayes and Sample Compression Hypernetworks (ICML 2025 · similarity 0.91)
- Does Compression Imply Generalization? A Minimum Description Length Perspective (NeurIPS 2026 · similarity 0.89)
- Generalizability of Neural Networks Minimizing Empirical Risk Based on Expressive Power (ICLR 2025 · similarity 0.87)
- Fantastic Generalization Measures are Nowhere to be Found (ICLR 2024 · similarity 0.87)
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.