Estimate Sequences for Variance-Reduced Stochastic Composite Optimization

Authors: Andrei Kulunchakov, Julien Mairal

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5. Experiments Following Bietti & Mairal (2017); Zheng & Kwok (2018) we consider logistic regression with Drop Out (Srivastava et al., 2014), which consists of randomly setting to zero each vector entry with probability δ, leading to the problem... We use three data sets alpha, ckn-cifar, and gene from different nature, which are presented in the appendix, along with other experimental details. Figure 1. Objective function value on a logarithmic scale with λ = 1/10n (left) and λ = 1/100n (right), with no Drop Out.
Researcher Affiliation Academia 1Univ. Grenoble Alpes, Inria, CNRS, Grenoble INP, LJK, 38000 Grenoble, France.
Pseudocode Yes Algorithm 1 Iteration (A) with random-SVRG estimator; Algorithm 2 Accelerated and robust random-SVRG
Open Source Code No The paper does not include an unambiguous statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes We consider logistic regression with Drop Out (Srivastava et al., 2014)... We use three data sets alpha, ckn-cifar, and gene from different nature, which are presented in the appendix, along with other experimental details.
Dataset Splits No The paper mentions using specific datasets but does not provide details on training, validation, or test splits (e.g., percentages or sample counts).
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments, such as GPU/CPU models or memory specifications.
Software Dependencies No The paper does not list specific software dependencies with version numbers that would be needed for reproducibility.
Experiment Setup Yes Following Bietti & Mairal (2017); Zheng & Kwok (2018) we consider logistic regression with Drop Out (Srivastava et al., 2014)... The parameter λ acts as a lower bound on µ and we consider λ = 1/10n, which is of the order of the smallest value that one would try when doing parameter search. We use three data sets alpha, ckn-cifar, and gene... We use them always with their theoretical step size, except rand-SVRG, which we evaluate with η = 1/3L in order to obtain a fair comparison with acc-SVRG. When using the decreasing step size strategy, we add the suffix -d to the method s name, and we consider also a minibatch variant of acc-SGD, denoted by acc-mb-SGD with minibatch size b = p L/µ. We also use the initial step size 1/3L for rand-SVRG-d since it performs better in practice.