Statistical Methods for ICT Applications


            Richa Vatsa

Richa Vatsa

Contact Details: School of Computer Science and Statistics,
Lloyd Institute (Room 108),
Trinity College,
Dublin 2,

+353 (0) 1 896 2048
+353 (0) 1 677 0711


    Summary of Research:

    My research work involes speeding up Bayesian computation using the Variational Bayes technique, a method of functional approximation that seeks to maintain accuracy. The Variational Bayes (VB) method provides a fair example of trade-off between accuracy and complexity of approximation. It may provide quick solutions to problems but is limited to the exponential family of distributions for successful application. 

    My work is focused on developing modifications in the VB algorithm and combining the method with other distributional approximational methods, for example, Laplace approximation or Gaussian approximation, so as to allow inference for complex problems in which more commonly used techniques are not applicable.

    I am currently working on the Variational Bayes inference for inverse regression problems. A motivating example of an inverse problem is palaeoclimate reconstruction. In the palaeoclimate reconstruction, past climate is inferred through pollen counts. A regression model of how pollen respond to change in climate is built and then inverted to infer ancient climate. Palaeoclimate reconstruction is a highly multivariate and computaionally intesive problem. This problem presents a complex and challenging model for the Variational Bayes method to accurately approximate such complexity.

    Conferences (Posters or Talks):
    1.  "VB approximation for Factor Analysis with parameterized factor loading matrix, and other examples of problems with its implementation'' at International Society for Bayesian Analysis (ISBA) 9th world meeting.

    2.  "The VB method with its application to complex problems eg. Palaeoclimate reconstruction, GLM" at 32nd Research Students'coneference (RSC).

    3.  "The Variational Bayes Method with its Application to Complex Problems eg. Palaeoclimate Reconstruction, GLM and Multimodality" at International Conference on Interdisciplinary  Mathematial and Statistical Techniques (IMST) 2009.

    4.  "The Variational Bayes Method for an Inverse Problem with Application to the Palaeoclimate Reconstruction" at 2009 Joint Statistical Meetings (JSM).

    5.  "The variational Bayes method for inverse problems with applications to the palaeoclimate reconstruction" at International Conference on Statistics, Probibility, Operation Research, Computer Science and alied Areas (IISA-ISPS) 2010.


    "R Vatsa and S.P. Wilson, The Variational Bayes Method For Inverse Regression Problems With an Application To The Palaeoclimate Reconstruction" a technical report 2009, TCD.