Big Data Analytics

With pervasive sensors continuously collecting and recording massive amounts of information, there is no doubt this is an era of data deluge. The sheer volume of data makes it impossible to run analytics on central processors and storage units. Large-scale datasets are often incomplete, prone to measurement corruption and communication errors, as well as vulnerable to cyber-attacks. Furthermore, as many data sources continuously generate data in real time, analytics must often be performed in real time, typically without a chance to revisit past entries. My ongoing research in this context aspires to tackle these challenges by putting forth a powerful model capturing a wide range of learning and processing tasks, and then offering architectures and algorithms to overcome the emerging practical hurdles, while revealing fundamental insights into the various analytical and implementation trade-offs involved. Recently, we proposed a new class of provable convergent, parallel decomposition algorithms to enable streaming analytics of sequential measurements using parallel processors. To the best of our knowledge, the proposed algorithmic framework is the first class of algorithms enabling parallel updates for nonconvex objective functions (without invoking any line-search). Other key features of the framework are: i) it uses random and/or greedy selection rules for the update of the block variables; i) it can use both first-order or high-order information; ii) if the solution of the optimization problem is expected to be sparse, it performs dimensionality reduction via active-set-based variable selection strategies; and iii) it includes as special cases several classical deterministic and random gradient-like schemes, such as (proximal) gradient or Newton type methods, block coordinate (parallel) descent schemes. The range of applicability of the developed framework is wide, including compressive sampling, matrix and tensor completion, outlier removal, distributed and online optimization.

Real-time Dynamic 3D Magnetic Resonance Imaging (MRI): Within the above context, a core application of my current research is devising novel encompassing real-time acquisition models applicable to different types of dynamic 3D MRI, as well as developing parallel algorithms for real-time MRI processing, implementable on multicore architectures. A first contribution on the topic has been putting forth, for the first time in MRI research, a low-rank-cognizant tensor -based model for free-breathing (rigid-body motion) cardiac (deformable motion) MRI acquisition, and consequent online high-quality recovery of dynamic images.

Related Publications

  • Amir Daneshmand, Francisco Facchinei, Vyacheslav Kungurtsev, and Gesualdo Scutari, "Hybrid Random/Deterministic Parallel Algorithms for Nonconvex Big Data Optimization," IEEE Trans. on Signal Processing, (submitted June 2014) [the order of the authors is alphabetic] [PDF].
  • Francisco Facchinei, Gesualdo Scutari, and Simone Sagratella, “Parallel Algorithms for Big Data Optimization,” IEEE Trans. on Signal Processing, to appear, 2015 [the first two authors contributed equally to the paper] [PDF].
  • Morteza Mardani, Leslie Ying, Gesualdo Scutari, Konstantinos Slavakis, and Georgios Giannakis, “Dynamic MRI Using Subspace Tensor Tracking,” in Proc. of the 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC’14), August 26-30, 2014, Chicago, IL, USA.
  • Gesualdo Scutari, Francisco Facchinei, Lorenzo Lampariello, and Peiran Song, “Distributed Methods for Nonconvex Constraints Multi-Agent Probles-Part I: Theory," IEEE Trans. on Signal Processing, (submitted Oct. 2014) [PDF] [Conference Version]
  • Gesualdo Scutari, Francisco Facchinei, Lorenzo Lampariello, and Peiran Song, “Distributed Methods for Nonconvex Constraints Multi-Agent Probles-Part II: Applications," IEEE Trans. on Signal Processing, (in preparation)