Stochastic Block BFGS: Squeezing More Curvature out of Data

Authors: Robert Gower, Donald Goldfarb, Peter Richtarik

ICML 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical tests on large-scale logistic regression problems reveal that our method is more robust and substantially outperforms current state-of-the-art methods.
Researcher Affiliation Academia Robert M. Gower GOWERROBERT@GMAIL.COM Donald Goldfarb GOLDFARB@COLUMBIA.EDU Peter Richt arik PETER.RICHTARIK@ED.AC.UK
Pseudocode Yes Algorithm 1 Stochastic Block BFGS Method; Algorithm 2 Block L-BFGS Update (Two-loop Recursion); Algorithm 3 Block L-BFGS Update (Factored loop recursion)
Open Source Code Yes All the code for the experiments can be downloaded from www.maths.ed.ac.uk/ prichtar/i software.html.
Open Datasets Yes We tested seven empirical risk minimization problems with a logistic loss and L2 regularizer using data from LIBSVM (Chang & Lin, 2011).
Dataset Splits No The paper does not explicitly provide specific training/validation/test dataset splits (e.g., percentages, sample counts, or explicit files).
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running experiments.
Software Dependencies No All the methods were implemented in MATLAB. (Does not specify version number for MATLAB or other dependencies).
Experiment Setup Yes We set the regularization parameter λ = 1/n for all experiments. We set the subsampling size |St| = n throughout our tests. We tested each method with a stepsize η {100, 5 · 10−1, 10−1, . . . , 10−6, 5 · 10−7, 10−7} for the best outcome, and used the resulting η. Finally, we used m = n/|St| for the number of inner iterations... We set the memory to 10 for the MNJ method in all tests...