Gradient Estimation with Simultaneous Perturbation and Compressive Sensing

Authors: Vivek S. Borkar, Vikranth R. Dwaracherla, Neeraja Sahasrabudhe

JMLR 2017 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We theoretically justify its computational advantages and illustrate them empirically by numerical experiments. In particular, applications to estimating gradient outer product matrix as well as standard optimization problems are illustrated via simulations. ... Section 3 presents applications to manifold learning and optimization with simulated examples. ... In this section, we test our method on real data.
Researcher Affiliation Academia Vivek S. Borkar EMAIL Department of Electrical Engineering Indian Institute of Technology Bombay Mumbai 400076, India; Vikranth R. Dwaracherla EMAIL Department of Electrical Engineering Stanford, USA; Neeraja Sahasrabudhe EMAIL Department of Mathematical Sciences Indian Institute of Science Education and Research, Mohali SAS Nagar 140306, India
Pseudocode Yes Algorithm 1 Gradient Estimation at some x Rn with SP and Compressive Sensing ... Algorithm 2 Stochastic Gradient Decent with Compressive Sensing
Open Source Code No The paper mentions a third-party toolbox used for simulations: "All the simulations were performed on MATLAB using the available toolbox for l1-minimization. (Berkeley database: http://www.eecs.berkeley.edu/ yang/software/ l1benchmark/)" but does not explicitly state that the authors' own code for the described methodology is released or available.
Open Datasets Yes Gradient estimation technique proposed in this paper is applied on UJIIndoor Loc Data Set (Torres-Sospedra et. al, 2014)
Dataset Splits Yes The Data-set has 19937 samples for training and 1111 samples for testing.
Hardware Specification No The paper mentions: "We thank GPU Centre of Excellence, IIT Bombay for providing us with the facility to carry out simulations" but does not specify any particular GPU model, CPU, or other detailed hardware specifications.
Software Dependencies No The paper states: "All the simulations were performed on MATLAB using the available toolbox for l1-minimization." However, no specific version numbers for MATLAB or the l1-minimization toolbox are provided.
Experiment Setup Yes Parameters like k (number of trials of SP) and iteration for l1 minimization are chosen so that both methods converge empirically in least number of iterations. Taking into account the higher error in SPSA, we take number of trials of SP to be 20 for Spall s SPSA and 10 for our method. Maximum number of iterations in Homotopy method are limited to 50.