Stochastic Frank-Wolfe for Composite Convex Minimization

Authors: Francesco Locatello, Alp Yurtsever, Olivier Fercoq, Volkan Cevher

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

Reproducibility Variable Result LLM Response
Research Type Experimental This section presents the empirical performance of the proposed method for the stochastic kmeans clustering, covariance matrix estimation, and matrix completion problems.
Researcher Affiliation Academia Francesco Locatello? Alp Yurtsever Olivier Fercoq Volkan Cevher francesco.locatello@inf.ethz.ch {alp.yurtsever,volkan.cevher}@epfl.ch olivier.fercoq@telecom-paristech.fr ?Department of Computer Science, ETH Zurich, Switzerland LIONS, Ecole Polytechnique F ed erale de Lausanne, Switzerland LTCI, T el ecom Paris, Universit e Paris-Saclay, France
Pseudocode Yes Algorithm 1 SHCGM Input: x1 2 X, β0 > 0, d0 = 0 for k = 1, 2, . . . , do k = 9/(k + 8) βk = β0/(k + 8) 1 2 k = 4/(k + 7) 2 3 dk = (1 k)dk 1 + krxf(xk, !k) vk = dk +β 1 Axk proxβkg(Axk) sk = arg minx2X xk+1 = xk + k(sk xk) end for
Open Source Code Yes We include the code to reproduce the results in the supplements.
Open Datasets Yes with a sample of 1000 datapoints from the MNIST data2. 2Y. Le Cun and C. Cortes. Available at http://yann.lecun.com/exdb/mnist/ and We consider a test setup with the Movie Lens100k dataset3 [13]. 3F.M. Harper, J.A. Konstan. Available at https://grouplens.org/datasets/movielens/ and We consider a test setup with the Movie Lens1m dataset3 with 1 million ratings from 6000 users on 4000 movies. 3F.M. Harper, J.A. Konstan. Available at https://grouplens.org/datasets/movielens/
Dataset Splits Yes We use the default ub.train and ub.test partitions provided with the original data. We partition the data into training and test samples with a 80/20 train/test split.
Hardware Specification Yes We performed the experiments in MATLAB R2018a using a computing system of 4 Intel Xeon CPU E5-2630 v3@2.40GHz and 16 GB RAM.
Software Dependencies Yes We performed the experiments in MATLAB R2018a using a computing system of 4 Intel Xeon CPU E5-2630 v3@2.40GHz and 16 GB RAM.
Experiment Setup Yes We use β0 = 1 for HCGM and β0 = 10 for SHCGM. We set these values by tuning both methods by trying β0 = 0.01, 0.1, ..., 1000. We set the model parameter for the nuclear norm constraint β1 = 7000, and the initial smoothing parameter β0 = 10. We set the model parameter β1 = 200000. We use β0 = 10 for SHCGM, and we set the step-size parameter γ = 1 for S3CCM.