Research map / Optimization / Nonconvex and smooth optimization
First-order Convergence Theory for Weakly-Convex-Weakly-Concave Min-max Problems
Mingrui Liu, Hassan Rafique, Qihang Lin, Tianbao Yang
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 extragradient instead of gradient descent as subproblem solver | IPP framework · Synthetic min-max, rho=10^-3 | gradient evaluations: 14857 -> 62 | Improved |
| use extragradient instead of gradient descent as subproblem solver | IPP framework · Synthetic min-max, rho=10^-7 | gradient evaluations: >10^8 -> 60 | Improved |
| use extragradient instead of gradient descent as subproblem solver | IPP framework · Synthetic min-max, rho=10^-1 | gradient evaluations: 9239 -> 1854 | Improved |
| use variance reduction (SVRG) instead of SGD for small target error | IPP framework · Synthetic finite-sum, eps=10^-10, rho=10^-1 | stochastic gradient evaluations: 12710 -> 1002 | Improved |
| use variance reduction (SVRG) instead of SGD for small target error | IPP framework · Synthetic finite-sum, eps=10^-10, rho=10^-3 | stochastic gradient evaluations: 2515 -> 1503 | Improved |
| use variance reduction (SVRG) instead of SGD for loose target error | IPP framework · Synthetic finite-sum, eps=10^-1, rho=10^-1 | stochastic gradient evaluations: 172 -> 1002 | Hurt |
| use variance reduction (SVRG) instead of SGD for loose target error | IPP framework · Synthetic finite-sum, eps=10^-1, rho=10^-3 | stochastic gradient evaluations: 67 -> 1025 | Hurt |
| wrap SGD in the inexact proximal point framework (IPP-SGD) | SGD · WGAN/WGAN-GP on CIFAR10 and LSUN Bedroom | inception score | Improved |
| wrap Adam in the inexact proximal point framework (IPP-Adam) | Adam · WGAN/WGAN-GP on CIFAR10 and LSUN Bedroom | inception score | Improved |
| use 5 discriminator steps per generator step (ablation) | Adam · WGAN training | inception score | Improved |
| use 5 discriminator steps per generator step (ablation) | SGD or WGAN-GP · WGAN-GP and SGD training | inception score | No clear effect |
| use Adam instead of SGD as inner solver in IPP | IPP · GAN training | convergence speed | Improved |
Closest papers by meaning
- Weaker MVI Condition: Extragradient Methods with Multi-Step Exploration (ICLR 2024 · similarity 0.90)
- Stability and Generalization of Stochastic Gradient Methods for Minimax Problems (ICML 2021 · similarity 0.90)
- A Convergent and Dimension-Independent Min-Max Optimization Algorithm (ICML 2022 · similarity 0.89)
- Optimistic Dual Extrapolation for Coherent Non-monotone Variational Inequalities (NeurIPS 2020 · similarity 0.89)
- Single-Loop Stochastic Algorithms for Difference of Max-Structured Weakly Convex Functions (NeurIPS 2024 · similarity 0.88)
- Delving into the Convergence of Generalized Smooth Minimax Optimization (ICML 2024 · similarity 0.87)
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.
