Research map / Learning theory and algorithms / Kernels and regression theory
The Pick-to-Learn Algorithm: Empowering Compression for Tight Generalization Bounds and Improved Post-training Performance
Dario Paccagnan, Marco Campi, Simone Garatti
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 P2L with GD instead of PAC-Bayes | GD classifier · MNIST 1000 samples delta=0.035 | bound on risk: 0.177 -> 0.143 | Improved |
| use P2L with GD instead of PAC-Bayes | GD classifier · MNIST 1000 samples delta=0.035 | test misclassification: 0.088 -> 0.072 | Improved |
| use P2L with GD instead of test-set approach | GD classifier · MNIST 1000 samples delta=0.035 | bound on risk: 0.129 -> 0.143 | Hurt |
| use P2L with GD instead of test-set approach | GD classifier · MNIST 1000 samples delta=0.001 | bound on risk: 0.155 -> 0.163 | Hurt |
| use P2L with GD instead of test-set approach | GD classifier · MNIST 1000 samples delta=0.035 | test misclassification: 0.079 -> 0.072 | Improved |
| use P2L instead of test-set approach (tightness) | GD classifier · MNIST 1000 samples delta=0.035 | difference bound minus test risk: 0.050 -> 0.071 | Hurt |
| use P2L instead of PAC-Bayes (running time) | GD classifier · MNIST | running time: 4m 1s -> 2m 1s | Improved |
| use P2L instead of test-set approach (running time) | GD classifier · MNIST | running time: 0m 5s -> 2m 1s | Hurt |
| vary pretrain portion of initial hypothesis h0 in P2L (ablation) | P2L · MNIST | test misclassification | No clear effect |
| use only 10% of data to pretrain h0 in P2L vs 60% prior for PAC-Bayes (ablation) | P2L · MNIST | bound and test performance | Mixed |
| use P2L instead of test-set approach | GD regressor · Synthetic regression N=200 | bound on risk: 0.250 -> 0.172 | Improved |
| use P2L instead of test-set approach | GD regressor · Synthetic regression N=200 | test risk: 0.173 -> 0.064 | Improved |
Closest papers by meaning
- Generalization Bounds via Meta-Learned Model Representations: PAC-Bayes and Sample Compression Hypernetworks (ICML 2025 · similarity 0.88)
- Compression, Generalization and Learning (JMLR 2023 · similarity 0.88)
- Does Compression Imply Generalization? A Minimum Description Length Perspective (NeurIPS 2026 · similarity 0.87)
- Certification-Enhanced Generalization Bounds (NeurIPS 2026 · similarity 0.86)
- Slicing Mutual Information Generalization Bounds for Neural Networks (ICML 2024 · similarity 0.86)
- A Universal Growth Rate for Learning with Smooth Surrogate Losses (NeurIPS 2024 · similarity 0.85)
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.
