This repository contains implementations for the numerical evaluations in the paper:
R. Saha, M. Pilanci and A. J. Goldsmith, "Efficient Randomized Subspace Embeddings for Distributed Optimization Under a Communication Budget," in IEEE Journal on Selected Areas in Information Theory, vol. 3, no. 2, pp. 183-196, June 2022, doi: 10.1109/JSAIT.2022.3198412. (https://ieeexplore.ieee.org/abstract/document/9857556)
The following are stand-alone scripts and directly running them should work.
-
compression_methods_map.m: Gives the result of Fig. 1(a). Comparison of different compression methods with and without near-democratic embedding.
-
DGD-DEF_comparison.m: Gives the result of Fig. 1(b). Variation of empirical convergence rate of DGD-DEF with bit-budget per dimension (R).
-
wall_clock_time_comparsion.m: Gives the result of Fig. 1(c). Wall clock times for computing near-democratic vs. democratic representations.
For simulations with a Support Vector Machine (SVM) on synthetic data,
-
Firstly, run generate_SVM_data.m to generate synthetic data for classification.
-
Then, run SVM_simulations.m to train an SVM classifier on the generated synthetic data and obtain plots in Figs. 2(a) and 2(b).
For classification on MNIST dataset using SVMs,
- Run SVM_simulations_MNIST.m as a standalone script.
We would appreciate if you reach out and report any issues.