Stochastic Halpern Iteration with Variance Reduction for Stochastic Monotone Inclusions

Authors: Xufeng Cai, Chaobing Song, Cristóbal Guzmán, Jelena Diakonikolas

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

Reproducibility Variable Result LLM Response
Research Type Experimental We now illustrate the empirical performance of stochastic Halpern iteration on robust least square problems. Specifically, given data matrix A ∈ R n×d and noisy observation vector b ∈ R n subject to bounded deterministic perturbation δ with ‖δ‖ ∞, robust least square (RLS) minimizes the worst-case residue as minx∈R d maxδ:‖δ‖∞≤ε ‖Ax − y‖ 2 2 with y = b + δ [16]. We consider solving MI induced from RLS with Lagrangian relaxation where u = (x, y)T and F(u) =
Researcher Affiliation Academia Xufeng Cai Department of Computer Sciences University of Wisconsin-Madison xcai74@wisc.edu Chaobing Song Department of Computer Sciences University of Wisconsin-Madison songcb16@gmail.com Cristóbal Guzmán Institute for Mathematical and Computational Eng. Facultad de Matemáticas and Escuela de Ingeniería Pontificia Universidad Católica de Chile crguzmanp@mat.uc.cl Jelena Diakonikolas Department of Computer Sciences University of Wisconsin-Madison jelena@cs.wisc.edu
Pseudocode Yes Algorithm 1: Stochastic Halpern-Cocoercive (Halpern) Algorithm 2: Extrapolated Stochastic Halpern-Monotone (E-Halpern) Algorithm 3: Restarted Extrapolated Stochastic Halpern-Sharp (Restarted E-Halpern)
Open Source Code Yes 2Code is available at https://github.com/zephyr-cai/Halpern.
Open Datasets Yes We use a real-world superconductivity dataset [22] from UCI Machine Learning Repository [15] for our experiment, which is of size 21263 × 81. To ensure the problem is concave in y, we need that λ > 1; in the experiments, we set λ = 1.5. For the experiment, we compare Halpern, E-Halpern, and Restarted E-Halpern algorithms with gradient descent-ascent (GDA), extragradient (EG) [28], and Popov’s method [42] in stochastic settings.
Dataset Splits No The main text describes the dataset but does not provide specific training/test/validation split percentages, sample counts, or detailed splitting methodology.
Hardware Specification Yes We implement all the algorithms in Python and run each algorithm using one CPU core on a mac OS machine with Intel 2.3GHz Dual Core i5 Processor and 8GB RAM.
Software Dependencies No The paper mentions implementation in 'Python' but does not specify its version or any other software dependencies with version numbers.
Experiment Setup Yes We use constant batch sizes and constant step sizes for GDA, EG, and Popov. We also choose the batch sizes of PAGE estimator to ensure E[‖F(uk) − e F(uk)‖^2] = O(1/k), which handles error accumulation [29] and early stagnation of stochastic Halpern iteration. We implement all the algorithms in Python and run each algorithm using one CPU core on a mac OS machine with Intel 2.3GHz Dual Core i5 Processor and 8GB RAM.2