Research map / Learning theory and algorithms / Kernels and regression theory
Tighter Risk Certificates for Neural Networks
María Pérez-Ortiz, Omar Rivasplata, John Shawe-Taylor, Csaba Szepesvári
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 |
|---|---|---|---|
| train with f_quad instead of f_classic | PBB FCN, Rand.Init. prior · MNIST | risk certificate (0-1 loss): .3304 -> .3155 | Improved |
| train with f_quad instead of f_classic | PBB CNN, Learnt prior · MNIST | risk certificate (0-1 loss): .0166 -> .0155 | Improved |
| train with f_quad instead of f_classic | PBB FCN, Learnt prior · MNIST | stochastic predictor test 0-1 error: .0230 -> .0202 | Improved |
| train with f_lambda instead of f_quad | PBB CNN, Rand.Init. prior · MNIST | stochastic predictor test 0-1 error: .0513 -> .0397 | Improved |
| train with f_lambda instead of f_quad | PBB FCN, Learnt prior · MNIST | risk certificate (0-1 loss): .0279 -> .0354 | Hurt |
| train with f_bbb (KL coefficient 0.1) instead of f_quad | PBB FCN, Rand.Init. prior · MNIST | stochastic predictor test 0-1 error: .0921 -> .0293 | Mixed |
| use data-dependent (Learnt) prior instead of data-free random prior (ablation) | f_quad FCN · MNIST | risk certificate (0-1 loss): .3155 -> .0279 | Improved |
| use data-dependent (Learnt) prior instead of data-free random prior (ablation) | f_quad CNN · MNIST | stochastic predictor test 0-1 error: .0513 -> .0104 | Improved |
| use CNN instead of FCN (ablation) | f_quad, Learnt prior · MNIST | risk certificate (0-1 loss): .0279 -> .0155 | Improved |
| sample stochastic predictor from posterior instead of prior mean predictor (ablation) | f_quad FCN, Rand.Init. prior · MNIST | test 0-1 error: .8792 -> .0921 | Improved |
| add KL attenuating coefficient 0.0001 to training objective (ablation) | f_quad CNN, Rand.Init. prior · MNIST | stochastic predictor test 0-1 error: .0513 -> .0305 | Mixed |
| train with PBB f_quad and SGD instead of SGLD (D&R 2018) | FCN with data-dependent prior · MNIST | risk certificate (0-1 loss): 0.2100 -> 0.0279 | Improved |
Closest papers by meaning
- Controlling Multiple Errors Simultaneously with a PAC-Bayes Bound (NeurIPS 2024 · similarity 0.87)
- Learning via Surrogate PAC-Bayes (NeurIPS 2024 · similarity 0.85)
- Recursive PAC-Bayes: A Frequentist Approach to Sequential Prior Updates with No Information Loss (NeurIPS 2024 · similarity 0.85)
- Sharper Guarantees for Learning Neural Network Classifiers with Gradient Methods (ICLR 2025 · similarity 0.85)
- Solving Probabilistic Verification Problems of Neural Networks using Branch and Bound (ICML 2025 · similarity 0.85)
- PAC-Bayes Generalisation Bounds for Dynamical Systems including Stable RNNs (AAAI 2024 · similarity 0.84)
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.
