Research map / Optimization / Nonconvex and smooth optimization
SEGA: Variance Reduction via Gradient Sketching
Filip Hanzely, Konstantin Mishchenko, Peter Richtarik
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 |
|---|---|---|---|
| Replace coordinate sketches with sketches from the correct gradient subspace (ablation) | SEGA · Constrained least squares with n much greater than d | convergence speed: roughly by the factor \tfrac{n}{d} | Improved |
| Build a persistent gradient estimate through sketch-and-project updates | SEGA · Composite optimization with non-separable regularizers | convergence | Improved |
| Use Gaussian gradient sketches rather than reconstructing the full gradient before taking a step | SEGA versus PGD · Synthetic quadratic problems with an l2-ball constraint | convergence speed | Improved |
| Estimate sketched gradients with finite differences and use SEGA rather than random direct search | Zeroth-order SEGA · Synthetic quadratic problems, Gaussian and coordinate directions | convergence | Improved |
| Accelerate SEGA using Nesterov-style updates | SEGA · Coordinate sketches | iteration complexity: 1.62 ... versus 9.8 ... | Mixed |
Closest papers by meaning
- Newton-LESS: Sparsification without Trade-offs for the Sketched Newton Update (NeurIPS 2021 · similarity 0.87)
- RSN: Randomized Subspace Newton (NeurIPS 2019 · similarity 0.87)
- Variance Reduced Coordinate Descent with Acceleration: New Method With a Surprising Application to Finite-Sum Problems (ICML 2020 · similarity 0.87)
- Gradient Projection Iterative Sketch for Large-Scale Constrained Least-Squares (ICML 2017 · similarity 0.87)
- Fundamental Bias in Inverting Random Sampling Matrices with Application to Sub-sampled Newton (ICML 2025 · similarity 0.86)
- Can Gaussian Sketching Converge Faster on a Preconditioned Landscape? (ICML 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.
