Research map / Learning theory and algorithms / Matrix and tensor methods
Near-Optimal Smoothing of Structured Conditional Probability Matrices
Moein Falahatgar, Mesrob I Ohannessian, Alon Orlitsky
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 |
|---|---|---|---|
| add low-rank structure with absolute-discounting smoothing instead of Kneser-Ney | bigram model · tartuffe | cross-entropy: 5.7555 -> 5.6923 | Improved |
| add low-rank structure with absolute-discounting smoothing instead of Kneser-Ney | bigram model · genesis | cross-entropy: 5.7341 -> 5.6673 | Improved |
| add low-rank structure with absolute-discounting smoothing instead of Kneser-Ney | bigram model · brown | cross-entropy: 7.7001 -> 7.609 | Improved |
| low-rank smoothing instead of Stupid Backoff | bigram model · tartuffe | cross-entropy: 6.0426 -> 5.6923 | Improved |
| low-rank smoothing instead of plain absolute discounting | bigram model · brown | cross-entropy: 7.9819 -> 7.609 | Improved |
| naive Add-1/2 smoothing then low-rank factorization vs Add-1/2 without structure (ablation) | low-rank bigram · synthetic uniform B rows | KL loss | Hurt |
| smooth counts inside the low-rank procedure instead of naive smoothing then factorizing (ablation) | low-rank bigram · synthetic | KL loss | Improved |
| Kneser-Ney on uniform-row synthetic data | Kneser-Ney · synthetic uniform B rows | KL loss | Hurt |
| low-rank smoothing vs Stupid Backoff at 15k training words | Smoothed Low Rank · tartuffe | cross-entropy diff: 0.1 nats/word | Improved |
Closest papers by meaning
- Dimension-free Concentration Bounds on Hankel Matrices for Spectral Learning (JMLR 2016 · similarity 0.86)
- Variational Smoothing in Recurrent Neural Network Language Models (ICLR 2019 · similarity 0.86)
- Matrix Product States for Inference in Discrete Probabilistic Models (JMLR 2021 · similarity 0.85)
- Loss factorization, weakly supervised learning and label noise robustness (ICML 2016 · similarity 0.85)
- Gradient Span Algorithms Make Predictable Progress in High Dimension (NeurIPS 2026 · similarity 0.84)
- Uncertainty Herding: One Active Learning Method for All Label Budgets (ICLR 2025 · 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.
