Fundamental Benefit of Alternating Updates in Minimax Optimization

Authors: Jaewook Lee, Hanseul Cho, Chulhee Yun

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 7. Experiments and subsequent subsections describe empirical evaluation. For instance, We conducted experiments on a (3 + 3)-dimensional SCSC quadratic game to visually compare the convergence speed of the algorithms in Figure 1. and We run further experiments on (100 + 100)-dimensional SCSC quadratic games to extensively compare GDA, EG, OGD, and Alex-GDA.
Researcher Affiliation Academia 1KAIST AI, South Korea.
Pseudocode Yes Algorithm 1 Sim-GDA and Alt-GDA and Algorithm 2 Alternating-Extrapolation GDA (Alex-GDA).
Open Source Code Yes The experiments in Tables 1 and 2 can be reproduced with our code available at Git Hub.3 github.com/Hanseul Jo/Alex-GDA
Open Datasets Yes For MNIST (Deng, 2012)... For CIFAR-10 (Krizhevsky et al., 2009)... For LSUN-Bedroom 64 64 dataset (Yu et al., 2015)...
Dataset Splits No The paper describes using MNIST, CIFAR-10, and LSUN Bedroom datasets but does not explicitly provide details about train/validation/test splits, only mentioning training for GANs.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models or memory specifications used for experiments.
Software Dependencies No The paper mentions using PyTorch and Adam optimizer but does not specify their version numbers or other software dependencies with version information.
Experiment Setup Yes Parameter tuning. We tuned step sizes and other parameters (like γ and δ of Alex-GDA) by grid search... We tune the momentum parameters mx, my { 0.99, 0.95, 0.9, 0.8, 0.7, ..., 0.9, 0.95, 0.99}. We tune γ and δ for Alex-GDA as γ, δ {0.5, 0.6, 0.7, ..., 3.0}.