RECAPP: Crafting a More Efficient Catalyst for Convex Optimization

Authors: Yair Carmon, Arun Jambulapati, Yujia Jin, Aaron Sidford

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

Reproducibility Variable Result LLM Response
Research Type Experimental Preliminary experiments on logistic regression problem indicate that our method is competitive with Catalyst-SVRG in practice.2
Researcher Affiliation Academia 1Tel Aviv University 2Stanford University.
Pseudocode Yes Algorithm 1: RECAPP
Open Source Code Yes Code available at: github.com/yaircarmon/recapp.
Open Datasets Yes We consider logistic regression on three datasets from lib SVM (lib): covertype (n = 581, 012, d = 54), real-sim (n = 72, 309, d = 20, 958), and a9a (n = 32, 561, d = 123). For each dataset we rescale the feature vectors to using unit Euclidean norm so that each fi is exactly 0.25-smooth. ... The LIBSVM data webpage. URL https://www. csie.ntu.edu.tw/ cjlin/libsvmtools/ datasets/.
Dataset Splits No The paper describes dataset usage but does not provide specific training, validation, or test splits. It only states that feature vectors were rescaled and no ℓ2 regularization was added.
Hardware Specification No The paper does not specify any particular hardware (e.g., GPU/CPU models, memory, or cloud instances) used for the experiments.
Software Dependencies No The paper mentions "Python" and the "Numba" package, but it does not specify any version numbers for these software components.
Experiment Setup Yes For each dataset we rescale the feature vectors to using unit Euclidean norm so that each fi is exactly 0.25-smooth. We do not add ℓ2 regularization to the logistic regression objective. ... For RECAPP and Catalyst, we tune the proximal regularization parameter λ (called κ in (Lin et al., 2017)). For each problem and each algorithm, we test λ values of the form αL/n, where L = 0.25 is the objective smoothness, n is the dataset size and α in the set {0.001, 0.003, 0.01, 0.03, 0.1, 0.3, 1.0, 3.0, 10.0}. ... In Algorithm 2 we set the parameters j0 = 0 and we test p {0, 0.1, 0.25, 0.5}.