HP-CONCORD

Massively Parallel Graphical Model Structure Learning Library

HP-CONCORD (High-Performance CONCORD) is a highly-scalable distributed-memory implementation of the CONCORD-ISTA algorithm for sparse inverse covariance matrix estimation. It is implemented in C++ with MPI and OpenMP. The main bottleneck of HP-CONCORD is iterative sparse-dense matrix-matrix multiplication which is handled by the SpDM3 library.

Downloads

The code is available on BitBucket. Please cite our papers in the publication section if you use the library.
HP-CONCORD is also available as a pre-compiled NERSC module on the Edison supercomputer. For more details, please refer to the instructions in our BitBucket page.

Documents

12/08/17: HP-CONCORD poster for the Advances in Modeling and Learning Interactions from Complex Data NIPS Workshop 2017.

Publications

Communication-Avoiding Optimization Methods for Distributed Massive-Scale Sparse Inverse Covariance Estimation [Paper] [arXiv] [BibTeX]
Penporn Koanantakool, Alnur Ali, Ariful Azad, Aydın Buluç, Dmitriy Morozov, Leonid Oliker, Katherine Yelick, and Sang-Yun Oh;
21st International Conference on Artificial Intelligence and Statistics (AISTATS 2018).
Also ArXiv e-prints, October 2017.
Communication-Avoiding Parallel Sparse-Dense Matrix-Matrix Multiplication [Paper] [Slides] [BibTeX]
Penporn Koanantakool, Ariful Azad, Aydın Buluç, Dmitriy Morozov, Sang-Yun Oh, Leonid Oliker, and Katherine Yelick;
30th IEEE International Parallel & Distributed Processing Symposium (IPDPS 2016), Chicago, IL, USA, May 2016.
Optimization methods for sparse pseudo-likelihood graphical model selection. [Paper] [BibTeX]
Sang-Yun Oh, Onkar Dalal, Kshitij Khare, and Bala Rajaratnam;
In NIPS 27, pages 667–675. 2014.

Contact

penpornk (at) eecs (dot) berkeley (dot) edu

Back to Penporn's home page.