Research map / Learning theory and algorithms / Matrix and tensor methods
Fast Screening Rules for Optimal Design via Quadratic Lasso Reformulation
Guillaume Sagnol, Luc Pronzato
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 homotopy algorithm instead of coordinate descent (exact vs tol 1e-4 solution) | quadratic Lasso c-optimal design · MNIST c-optimal, lambda=10^-4 | time (s): 2012.15 -> 44.66 | Improved |
| use homotopy algorithm instead of coordinate descent | quadratic Lasso c-optimal design · MNIST c-optimal, lambda=10^0 | time (s): 1.68 -> 0.22 | Improved |
| use SOCP-2 direct reformulation instead of SOCP-1 | Gurobi SOCP solver · MNIST c-optimal, lambda=10^0 | time (s): 62.57 -> 13.54 | Improved |
| use MWU instead of CD at small lambda | first-order solvers with screening D1 · MNIST c-optimal, lambda=10^-4 | time (s): 2012.15 -> 619.09 | Improved |
| decrease lambda from 10^0 to 10^-4 for first-order CD (ablation) | CD with screening D1 · MNIST c-optimal | time (s): 1.68 -> 2012.15 | Hurt |
| decrease lambda from 10^0 to 10^-4 for SOCP-2 (ablation) | Gurobi SOCP-2 · MNIST c-optimal | time (s): 13.54 -> 13.91 | No clear effect |
| screen with D2 instead of D1 (more frequent screening) (ablation) | CD with screening · MNIST c-optimal, lambda=0.4 | convergence rate | Mixed |
| add screening rule D1 every 10 iterations to block CD (ablation) | block CD for L-optimal design · MNIST L-optimal (p=1200, m=400, r=50) | convergence speed: acceleration factor of about two | Improved |
| add screening rule D1 to multiplicative weight update (ablation) | MWU for L-optimal design · MNIST L-optimal (p=1200, m=400, r=50) | convergence speed | Improved |
Closest papers by meaning
- Experimental Design on a Budget for Sparse Linear Models and Applications (ICML 2016 · similarity 0.90)
- The Hessian Screening Rule (NeurIPS 2022 · similarity 0.89)
- A Practical Scheme and Fast Algorithm to Tune the Lasso With Optimality Guarantees (JMLR 2016 · similarity 0.89)
- Screening Tests for Lasso Problems (TPAMI 2016 · similarity 0.89)
- Screening rules for Lasso with non-convex Sparse Regularizers (ICML 2019 · similarity 0.88)
- L1-Regularized Least Squares for Support Recovery of High Dimensional Single Index Models with Gaussian Designs (JMLR 2016 · 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.
