Greedy and Random Quasi-Newton Methods with Faster Explicit Superlinear Convergence

Authors: Dachao Lin, Haishan Ye, Zhihua Zhang

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we verify our theorems through numerical results for quasi-Newton methods. Rodomanov and Nesterov [19] have already compared their proposed greedy quasi-Newton method with the classical quasi-Newton methods.
Researcher Affiliation Collaboration Dachao Lin1 Haishan Ye2 Zhihua Zhang3 1Academy for Advanced Interdisciplinary Studies, Peking University 2School of Management, Xi an Jiaotong University; Shenzhen Research Institute of Big Data 3School of Mathematical Sciences, Peking University; Huawei Noah s Ark Lab
Pseudocode Yes Algorithm 1 Greedy/Random SR1 update; Algorithm 2 Greedy/Random BFGS update; Algorithm 3 Greedy/Random SR1/BFGS methods for quadratic minimization; Algorithm 4 Greedy/Random SR1/BFGS methods for strongly self-concordant objective
Open Source Code No The paper does not provide any statements or links indicating that its source code is open-source or publicly available.
Open Datasets Yes We follow the same experimental design but take data from the LIBSVM collection of real-world data sets for binary classification problems [6].
Dataset Splits No The paper mentions using datasets for experiments but does not provide specific details on how the data was split into training, validation, and test sets (e.g., percentages, sample counts, or cross-validation setup).
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., CPU, GPU models, memory) used to run the experiments.
Software Dependencies No The paper does not specify any software dependencies with version numbers (e.g., libraries, frameworks, or programming language versions) used for the experiments.
Experiment Setup Yes In order to simulate the local convergence, we use the same initialization after running several standard Newton s steps to make measure f(w0) small (around 10 2 101). And we do not apply the correction strategy ( Gk = (1 + Mrk)Gk) in Algorithm 4).