Research map / Learning theory and algorithms / Kernels and regression theory
PAC-Bayes Compression Bounds So Tight That They Can Explain Generalization
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 |
|---|---|---|---|
| Train and quantize in an intrinsic-dimension subspace instead of pruning (ablation) | PAC-Bayes compression bounds · MNIST, FashionMNIST, SVHN, CIFAR-10, CIFAR-100 | PAC-Bayes error bound (%): 47.9 -> 11.6; 54.9 -> 32.8; 74.4 -> 36.1; 100.0 -> 58.2; 99.9 -> 94.6 | Helped |
| Use ImageNet pretraining for downstream transfer | Subspace compression bounds · CIFAR-10 | PAC-Bayes error bound (%): 58.2 -> 35.1 | Helped |
| Use ImageNet pretraining for downstream transfer | Subspace compression bounds · CIFAR-100 | PAC-Bayes error bound (%): 94.6 -> 81.3 | Helped |
| Use ImageNet pretraining for downstream transfer | Subspace compression bounds · SVHN | PAC-Bayes error bound (%): 36.1 -> 29.1 | Helped |
| Replace stochastic training with full-batch training | LeNet-5 · MNIST | PAC-Bayes error bound (%): 11.6 -> 11.2 | No clear effect |
| Replace stochastic training with full-batch training | ResNet-18 · CIFAR-10 | PAC-Bayes error bound (%): 74.7 -> 75.3 | No clear effect |
| Replace stochastic training with full-batch training | LeNet-5 transfer model · SVHN to MNIST | PAC-Bayes error bound (%): 9.0 -> 8.7 | No clear effect |
| Use rotational equivariance | WideResNet · RotMNIST | PAC-Bayes error bound | Helped |
| Use rotational equivariance | WideResNet · MNIST | PAC-Bayes error bound | No clear effect |
| Shuffle image pixels | CNN on CIFAR-10 · CIFAR-10 with shuffled pixels | compressibility | Hurt |
| Shuffle labels | CNN and MLP · CIFAR-10 with shuffled labels | PAC-Bayes error bound (%): 100% | Hurt |
| Use a CNN instead of a same-parameter-count MLP | Image classifier · CIFAR-10 | PAC-Bayes error bound | Helped |
Closest papers by meaning
- Generalization Bounds via Meta-Learned Model Representations: PAC-Bayes and Sample Compression Hypernetworks (ICML 2025 · similarity 0.92)
- Slicing Mutual Information Generalization Bounds for Neural Networks (ICML 2024 · similarity 0.89)
- Does Compression Imply Generalization? A Minimum Description Length Perspective (NeurIPS 2026 · similarity 0.88)
- Tighter CMI-Based Generalization Bounds via Stochastic Projection and Quantization (NeurIPS 2025 · similarity 0.88)
- Fantastic Generalization Measures are Nowhere to be Found (ICLR 2024 · similarity 0.88)
- Generalizability of Neural Networks Minimizing Empirical Risk Based on Expressive Power (ICLR 2025 · similarity 0.88)
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.