Research map / Optimization / Distributed and compressed
A distributed block coordinate descent method for training l1 regularized linear classifiers
Dhruv Mahajan, S. Sathiya Keerthi, S. Sundararajan
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 greedy S-scheme instead of random R-scheme variable selection (ablation) | DBCD · KDD (P=25, tau=-1) | outer iterations: 236 -> 8 | Improved |
| use greedy S-scheme instead of random R-scheme variable selection (ablation) | PCD · KDD (P=25, tau=-1) | outer iterations: 294 -> 12 | Improved |
| use greedy S-scheme instead of random R-scheme variable selection (ablation) | DBCD · URL (P=25, tau=0) | outer iterations: 101 -> 6 | Improved |
| use greedy S-scheme instead of random R-scheme variable selection (ablation) | PCD · URL (P=25, tau=-1) | outer iterations: 796 -> 196 | Improved |
| replace decoupled quadratic with joint nonlinear block optimization (DBCD-S vs PCD-S) (ablation) | S-scheme methods · URL (P=25, tau=0) | outer iterations: 65 -> 6 | Improved |
| replace decoupled quadratic with joint nonlinear block optimization (DBCD-S vs PCD-S) (ablation) | S-scheme methods · KDD (P=25, tau=-2) | outer iterations: 311 -> 104 | Improved |
| replace decoupled quadratic with joint nonlinear block optimization (DBCD-S vs PCD-S) (ablation) | S-scheme methods · KDD (P=100, tau=-3) | outer iterations: 650 -> 668 | Hurt |
| add line search to Richtarik-style updates (PCD-R) (ablation) | RICHTARIK · KDD (P=25, WSS r=0.1) | computation time per iteration (s): 0.022 -> 0.138 | Hurt |
| add 10 inner CD cycles and S-scheme to block update (DBCD-S vs DBCD-R) (ablation) | DBCD · KDD (P=25, WSS r=0.1) | computation time per iteration (s): 0.991 -> 5.054 | Hurt |
| increase nodes from 25 to 100 | DBCD-S · KDD (tau=-2) | outer iterations: 104 -> 137 | No clear effect |
| increase nodes from 25 to 100 | DBCD-S · URL (tau=-1) | outer iterations: 22 -> 28 | No clear effect |
| use ADMM instead of RICHTARIK | distributed l1 solver · URL (P=25, tau=-1) | outer iterations: >800 -> 214 | Improved |
Closest papers by meaning
- An efficient distributed learning algorithm based on effective local functional approximations (JMLR 2018 · similarity 0.96)
- CoCoA: A General Framework for Communication-Efficient Distributed Optimization (JMLR 2018 · similarity 0.87)
- Distributed Learning with Regularized Least Squares (JMLR 2017 · similarity 0.87)
- DSCOVR: Randomized Primal-Dual Block Coordinate Algorithms for Asynchronous Distributed Optimization (JMLR 2019 · similarity 0.86)
- Learning Sparse Classifiers: Continuous and Mixed Integer Optimization Perspectives (JMLR 2021 · similarity 0.86)
- A General Distributed Dual Coordinate Optimization Framework for Regularized Loss Minimization (JMLR 2017 · similarity 0.86)
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.
