Quasi-Newton Methods for Saddle Point Problems

Authors: Chengchang Liu, Luo Luo

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our numerical experiments show proposed algorithms outperform classical first-order methods. and Section 5 Numerical Experiments with subsections like AUC Maximization and Adversarial Debiasing and Figure 1 and 2 showing iteration numbers vs. g(z) 2 and CPU time (second) vs. g(z) 2.
Researcher Affiliation Academia Chengchang Liu Department of Computer Science & Engineering The Chinese University of Hong Kong 7liuchengchang@gmail.com Luo Luo School of Data Science Fudan University luoluo@fudan.edu.cn
Pseudocode Yes Algorithm 1 Fast-Chol(H, L, u), Algorithm 2 Random-Broyden-Quadratic, Algorithm 3 Random-BFGS-Quadratic, Algorithm 4 Random-SR1-Quadratic, Algorithm 5 Random-Broyden-General, Algorithm 6 Random-BFGS-General, Algorithm 7 Random-SR1-General
Open Source Code Yes Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] We present them in the supplemental material.
Open Datasets Yes We evaluate all algorithms on three real-world datasets a9a , w8a and sido0 . and fairness-aware binary classification dataset adult , bank market and law school [30].
Dataset Splits No The provided text mentions using datasets but does not explicitly detail training, validation, and test splits with percentages or specific methods for partitioning the data.
Hardware Specification Yes Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See Appendix E.
Software Dependencies No The paper does not provide specific version numbers for software dependencies such as libraries or programming languages used in the experiments.
Experiment Setup Yes We set λ = 100/n (for AUC Maximization) and We set the parameters β, λ and γ as 0.5, 10 4 and 10 4 respectively. The dimension of the problem is d = m + 1. Since the objective function is non-quadratic, we conduct the proposed algorithms in Section 3.3 (Algorithm 5, 6 and 7) here. We use extragradient as warm up to achieve the local condition for proposed algorithms. (for Adversarial Debiasing).