Generalization Properties and Implicit Regularization for Multiple Passes SGM

Authors: Junhong Lin, Raffaello Camoriano, Lorenzo Rosasco

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

Reproducibility Variable Result LLM Response
Research Type Experimental We carry out some numerical simulations to illustrate our results. The experiments are executed 10 times each, on the benchmark datasets reported in Table 1
Researcher Affiliation Academia Junhong Lin JHLIN5@HOTMAIL.COM Raffaello Camoriano , , RAFFAELLO.CAMORIANO@IIT.IT Lorenzo Rosasco , LROSASCO@MIT.EDU LCSL, Massachusetts Institute of Technology and Istituto Italiano di Tecnologia, Cambridge, MA 02139, USA DIBRIS, Universit a degli Studi di Genova, Via Dodecaneso 35, Genova, Italy i Cub Facility, Istituto Italiano di Tecnologia, Via Morego 30, Genova, Italy
Pseudocode Yes Algorithm 1. Given a sample z, the stochastic gradient method (SGM) is defined by w1 = 0 and wt+1 = wt ηt V (yjt, wt, Φ(xjt) )Φ(xjt), t = 1, . . . , T, for a non-increasing sequence of step-sizes {ηt > 0}t N and a stopping rule T N.
Open Source Code Yes Code: lcsl.github.io/Multiple Passes SGM (Footnote 4)
Open Datasets Yes Datasets: archive.ics.uci.edu/ml and www.csie.ntu.edu.tw/~cjlin/libsvmtools/ datasets/ (Footnote 5)
Dataset Splits Yes In order to apply hold-out cross-validation, the training set is split in two parts: one for empirical risk minimization and the other for validation error computation (80% 20%, respectively).
Hardware Specification Yes The experimental platform is a server with 12 Intel Xeon E5-2620 v2 (2.10GHz) CPUs and 132 GB of RAM.
Software Dependencies No The paper mentions using LIBSVM and implementing in Python (via a code link), but it does not specify version numbers for any software libraries, frameworks, or tools used in the experiments.
Experiment Setup Yes In the first experiment, the SIGM step-size is fixed as η = 1/ n. In the second experiment, we consider SIGM with decaying stepsize (η = 1/4 and θ = 1/2).