SANJUKTA BHOWMICK


Assistant Professor
Department of Computer Science and Engineering
Pennsylvania State University
Phone: (814)-865-0994, E-mail: bhowmick@cse.psu.edu

 

RESEARCH

COURSES

PUBLICATIONS

RESUME

RESEARCH


    My interests are in high performance computing, specifically in areas that involve a liaison between non-numerical (graph theory, machine learning) and numerical (large scale sparse linear solvers) algorithms. Core areas of my investigation include combinatorial scientific computing and solution of large scale sparse linear systems. Current research projects include;
  • Multimethod Solvers
      Design of multimethod solvers involves the development of robust and scalable algorithms for sparse linear systems by leveraging the strength of existing linear solvers. We developed the composite multimethod solver where the linear system is applied to a sequence of solvers and the adaptive multimethod solver where linear solvers are dynamically selected to match evolving linear system properties. Related Publications
      (With: Dinesh Kaushik, Lois Curfman McInnes, Boyana Norris, Padma Raghavan)
  • Machine Learning for Linear Solvers
      Linear solvers are designed to cater to different solution requirements and their performance of these depend on the properties of the linear systems being solved. Matching linear solvers to problem characteristics can produce near optimal performance. However, given the wide variety of linear solvers, the algorithmic options explode combinatorially. We use machine learning techniques to select "good" solvers according to linear system characteristics. Related Publications
      (With: Victor Eijkhout ,Yoav Freund, Erika Fuentes, David Keyes)
  • Automatic Differentiation
      Automatic differentiation is based on evaluating the exact derivatives of functions by using the chain rule, in order to avoid the inevitable numerical errors of Taylor's method. Function sequences in the chain rule are represented as directed acyclic graphs, therefore, algorithms for automatic differentiation can be derived through combinatorial techniques. We are investigating the optimization Hessian derivatives calculation through detecting symmetry in the computational graph. Related Publications
      (With: Paul Hovland)
  • Applications of Graph Embedding
      Graph embedding is used to arrange vertices and edges in an aesthetically pleasing pattern on a two dimensional space. Extending this concept to placement in an n-dimensional space we use embedding algorithms to optimize data layout in caches and in other data clustering applications like text mining. We are also considering use of embedding algorithms for computational biology applications like modeling of chromatin fibers.
      (With: Anirban Chaterjee, Mahmut Kandemir, Padma Raghavan)

    • LINKS TO COURSES

    LINKS TO SELECTED PUBLICATIONS


    RESUME