 |
Post-doctoral fellow Email: mrinal dot das at aalto dot fi |
Moving
I am going to join Prof. Andrew McCallum's group at UMass, Amherst, USA from August, 2016.
Brief Bio
Currently, I am a post-doctoral fellow with
Prof. Samuel Kaski at Department of Computer Scienece, Aalto Univesity in Helsinki, Finland since October, 2014.
I have been working on building differentially private Bayesian models for personalized medicine.
I have done my PhD at Department of Computer Science and Automation, Indian Institute of Science, India under the guidance of
Prof. Chiranjib Bhattacharyya.
PhD Thesis: Extensions and Applications of Stick-Breaking Process on Topic Models.
Research Interests
My fascination is to develop simple and novel mathematical models to address interesting and challenging practical problems. Broadly, I am interested in
Topic models and Bayesian nonparametric models for practical problems.
Differentially private Bayesian models
I am proficient and experienced in using both variational approximation as well as MCMC based inference for topic models. I like to explore different areas of application and in recent past I have worked with various types of text datasets like software projects, speech transcripts, multi-lingual corpora, news/blogs and comments. I have observed some challenging problems associated with these applications and developed novel mathematical models to solve them.
Privacy has become important after digitization of each and every information which can be personal, private or public. My interest is to explore, if it is possible to utilize the provable privacy guarantees of differential privacy within Bayesian framework despite maintaining accuracy. Specially in the field of personalized medicine, privacy protection has become mandatory due to sensitive genomic information. Similar privacy issues are prevalent in social media, advertisements and recommendation systems.
Research Achievements
Topic models are popular mathematical tools for analysing text datasets, where a corpus is a collection of documents. The state of art notion in topic models was to use single topic vector per document. I conceived the novel yet simple idea of using multiple topic vectors (MTV). We have observed phenomenal ability of MTV in (i) discovering subtle topics, (ii) modeling specific correspondence, (iii) modeling multi-glyphic topical correspondence, (iv) content driven user profiling for comment-worthy recommendations. All of them helped in inventing novel models (i) subtle topic models (STM, in ICML, 2013), (ii) specific correspondence topic models (SCTM, in WSDM, 2014), (iii) multi-glyphic correspondence topic model (AAAI, 2015), (iv) collaborative correspondence topic models (CCTM in RecSys, 2015).
I have also worked on Bayesian nonparametric models for learning very large scale (more than 8 million documents and 700 million tokens) datasets. There is NO method known using MCMC for such scale without using expensive parallel hardware. The work has been accepted at ICML, 2015.
Recently at Aalto university, we have been able to solve the problem of predicting drug sensitivity using gene expressions even after preserving privacy. I have conceived the novel concept of projecting outliers to tighter bounds without affecting non-outliers which has been the key in our method.