Educational Background

I did my Undergraduate from IIT Madras from the Department of Electrical Engineering and graduated in the year 2005.

I did my PhD at Johns Hopkins (2011).

Starting Oct 2011, I will be working in Google based in Mountain View California.

My resume is available here.

Major Software Undertakings

    • Hidden Markov Model Toolkit ( link). The entire zip file (with examples) is available here.

    • Vector Autoregressive Model Toolkit ( link).

Apart from this, I also have written end to end codes for

    • Multiple Layer Neural Network training,

    • Language to Language transliteration (uses the Google FST for transducing substrings). Our work was evaluated in the top 10 in the ACL NEWS workshop for several language pairs,

    • Maximum Entropy training toolkit for Language Modeling



  • PhD Dissertation: Learning and Inference Algorithms for Dynamical System Models of Dextrous Motion (pdf).


    • Balakrishnan Varadarajan, Sanjeev Khudanpur and Trac. D. Tran, "Stepwise Optimal Subspace Pursuit for Improving Sparse Recovery", Signal Processing Letters, 2010 (pdf). Extended report (pdf)

    • Balakrishnan Varadarajan, Sivaram Garimella and Sanjeev Khudanpur, Dirichlet Mixture Models to model Neural Network Posteriors in the HMM framework, Accepted in ICASSP-2011

  • Balakrishnan Varadarajan and Sanjeev Khudanpur, Learning and Inference Algorithms for Partially observed Structured Switching Vector Autoregressive Models, Accepted in ICASSP-2011 (pdf)


    • Balakrishnan Varadarajan and Delip Rao É›-extension Hidden Markov Models and Weighted Transducers for Machine Transliteration , ACL - Workshop-2009 (pdf)

    • Balakrishnan Varadarajan, Dong Yu, Li Deng, Alex Acero, "Using Collective Information in Semi-Supervised Learning for Speech Recognition", ICASSP 2009, Taipei, Taiwan.

    • Balakrishnan Varadarajan, Dong Yu, Li Deng, Alex Acero, "Maximizing Global Entropy Reduction for Active Learning In Speech Recognition", ICASSP 2009, Taipei, Taiwan.

    • Dong Yu, Balakrishnan Varadarajan; Li Deng, Alex Acero, "Active Learning and Semi-supervised Learning for Speech Recognition: A Unified Framework using the Global Entropy Reduction Maximization Criterion", CSL08-92

    • Balakrishnan Varadarajan, Carol Reiley, Henry Lin, Sanjeev Khudanpur, Gregory Hager, 'Data-Derived Models for Segmentation with Application to Surgical Assessment and Training', MICCAI 2009, 426-434 (pdf)


    • Balakrishnan Varadarajan, Sanjeev Khudanpur and Emmanuel Dupoux, ' Unsupervised Learning of Acoustic Sub-word Units', ACL-2008:HLT(In proceedings). (pdf)

    • Balakrishnan Varadarajan and Sanjeev Khudanpur, 'Automatically Learnin Speaker-independent Acoustic Subword Units', Interspeech 2008

    • Balakrishnan Varadarajan ,Daniel Povey & Stephen M Chu, 'Quick FMLLR for Speaker Adaptation in Speech Recognition', ICASSP 2008

    • Daniel Povey, Stephen M Chu & Balakrishnan Varadarajan, 'Universal Background Model Based Speech Recognition', ICASSP 2008

    • C.E. Reiley, H.C. Lin, B. Varadarajan, S. Khudanpur, D. D. Yuh, and G. D. Hager, ' Automatic Recognition of Surgical Motions Using Statistical Modeling for Capturing Variability', MMVR 2008.

Submitted Papers

    • Data driven Statistical Models for Computer Integrated Surgery (PAMI) (pdf).

PDF versions of selected publications can be found here.

Select Talks

    • Learning and Inference Algorithms for Structured Switching Vector AR Models. Slides are available here

    • Hidden Markov Model Toolkit for coarsely segmented motions. Slides are here

Research and technical interests

    • Machine Learning: I am interested in exploring areas of Machine Learning that require advanced mathematical thinking and algorithmic tools. I am more inclined towards principled techniques rather than ad-hoc ones. The primary areas I have worked in include building computer models for Speech recognition and robot motions. In the last few years, I have devoloped algorithms to learn structure in smooth time-varying signals. I have worked on Linear dynamical system models for trajectory data.

    • Compressed Sensing: I have developed algorithms that recover sparse signals more efficiently compared to pre-existing greedy techniques. Compressed sensing has vast applications in image compression. An application of my interest is to apply sparse learning for linear dynamical systems and vector autoregressive models to predict the current sample using a compact representation of the previous samples.

    • Algorithms : Apart from research and publishing, I have also got passion for algorithms and programming. I have interest in Mathematics, Mathematical puzzles and puzzles that require smart programming. I regularly participate in timed programming contests, notable topcoder, codeforces. Apart from timed ones, I also have interest in untimed ones and enjoy solving very hard problems. SPOJ and codechef are some of my favorites. My profile in SPOJ and codechef are available here (link) and here (link) respectively. I have won several contests. I was ranked 163 in Google Codejam Round 2, 2011 (link). I am almost always ranked in the top 20 fastest solvers in recent problems in projecteuler (link).