Our group is performing research in large-scale distributed storage and information processing.
PhD students @ UT Austin
Matrix Factorizations and Sparse PCA
Multiple components for SPCA using bipartite matchings (pdf) NIPS 2015
Orthogonal Nonnegative Matrix Factorization through Subspace Exploration (pdf) NIPS 2015
Stay on path: PCA along graph paths (pdf)
Learning sparse polynomials and causality
Sparse polynomial learning and Graph sketching (pdf) NIPS 2014
Learning Causal graphs with Small Interventions (pdf) NIPS 2015
Big graph analytics
Estimating profiles and motifs of large graphs (pdf) KDD 2015.
Distributed estimation of Graph 4-profiles (pdf) WWW 2016.
New erasure codes for big data storage.
Xorbas HDFS is a module for Hadoop Mapreduce that implements Locally repairable codes. These new codes can perform rebuilds with significantly smaller disk IO and network traffic compared to Reed-Solomon and other other classic codes. Our prototype is open source and available on Github and was tested on Amazon and Facebook clusters. It is a modification of the HDFS RAID module used by Facebook for big data storage protection.
The main idea of this project is to add caching capabilities to mobile femtocell stations. Several algorithmic and coding questions arise.
Connecting coding theory and compressed sensing through LP relaxations
We show that the standard LP relaxations for binary channel coding and compressed sensing have a non-trivial connection. This allows us to leverage results from LDPC codes to establish performance guarantees for basis pursuit in compressed sensing.