177 Huntington Avenue
Boston, MA 02115
- PhD, Indian Institute of Science – Bangalore, India
Sahely Bhadra is a postdoctoral researcher in the Khoury College of Computer Sciences working with Professor Tina Eliassi-Rad in the algorithms and theory program. Prior to joining Northeastern, Sahely worked on multi-view kernel completion and principal metabolic flux mode analysis as a postdoctoral researcher at Helsinki Institute of Information Technology. Sahely also worked as a postdoctoral researcher at Max Planck Institute of Informatics focusing on the correction of noisy labels via mutual consistency check. Sahely earned her PhD from the Indian Institute of Science in Bangalore, India.
- Hometown: Kolkata, India
- Field of Research/Teaching: Machine Learning and Optimization
What are the specific areas of your graduate education?
I have a PhD in Machine Learning and Computer Science along with IBM PhD Fellowship Award. Some of my professional positions include working a postdoctoral researcher at Helsinki Institute of Information Technology (HIIT) from October 2014 to December 2016. I worked on multi-view kernel completion and principal metabolic flux mode analysis with Professor Samuel Kaski and Professor Juho Rousu. I also worked as a postdoctoral researcher at Max Planck Institute of Informatics (MPII) from September 2012 to August 2014. I worked on the correction of noisy labels via mutual consistency check with Prof. Matthias Hein of Machine Learning Lab at Saarland University.
What are your research interests?
I am experienced in developing novel mathematical models to address interesting and challenging practical problems. Broadly, I have worked on methods such as kernel methods, convex and non-convex optimization, robust learning, and data analysis (PCA, CCA etc) with structure based regularization. I have also researched noisy features and annotation, handling missing values, and considering structural constraints for analyzing biological data. I like teaching and researching, so I had this academic path in mind.
What’s one problem you’d like to solve with your research/work?
I would like to count the uncountable part of real world data and extract useful information from unstructured noisy data.
What aspect of what you do is most interesting?
Most importantly, I extract important information from noisy data to figure out how to use information in incomplete data. It is interesting to me that I can bridge machine learning with robust optimization technique.
Some of my research achievements include:
- Multi-view Kernel Completion: We introduced the first method for completing kernel matrices that contain missing rows and columns as opposed to individual missing kernel values, with the help of information from other incomplete kernel matrices. This method does not require any of the kernels to be complete, and it can tackle non-linear kernels. Kernel completion is done by finding an appropriate set of related kernels for each missing entry from the set of available incomplete kernels.
- Principal Metabolic Flux Mode Analysis – We introduced structural PCA for data analysis and proposed a novel principal component analysis method regularized by stoichiometric information to discover underlying principal fluxes from given gene expression data. More importantly, this method does not require explicit knowledge of pathways or elementary flux modes for the system.
- Correction of Noisy Labels via Mutual Consistency Check: We proposed a new pre-learning method to correct noisy annotations using minimal human effort. Moreover, we proposed a novel algorithm using the Spannogram framework which solves the resultant non-convex problem efficiently with provable approximation guarantee.
- Classification with Uncertain Observation and Robust Optimization: During my PhD, I worked on building robust classifiers to deal with interval valued uncertainty in data as well as kernel matrices using robust optimization technique. I also dealt with scalability and robustness of these classifiers. We proposed a novel distribution free of large deviation inequality which handles uncertainty as well as a mirror descent algorithm (MDA) like procedures to scale the knowledge of proposed robust formulations.
What are your research or career goals, going forward?
I want to continue in academia and begin teaching.