Research summary
** I do not update this page regularly, please check my latest CV for more recent works and projects.
At IBM Research, I worked on many exciting problems over the last two years, including neural network verification; individual discrimination in tabular machine learning models; testing speech-to-text models for robustness and fairness; optimal decision trees and decision tree surrogates; model comparison; program synthesis, and API testing. These projects led to many client POCs and demos, 3 paper submissions, and 6 filed patents.
At IBM Research, I worked on many exciting problems over the last two years, including neural network verification; individual discrimination in tabular machine learning models; testing speech-to-text models for robustness and fairness; optimal decision trees and decision tree surrogates; model comparison; program synthesis, and API testing. These projects led to many client POCs and demos, 3 paper submissions, and 6 filed patents.
I gained knowledge on different sub-fields of machine learning, namely explainability methods and their drawbacks, fairness violations in machine learning models, and sparsity and other challenges in interpretable machine learning. These areas also helped me grasp numerous mathematical tools in probabilistic modeling and optimization, and shaped my research goals. I hope to use these skills to pursue research in trustworthy machine learning in near future.
Research internships
In the summer of 2019, I interned at Prof. Amit Sethi's lab at IIT Bombay. There I studied prostate cancer recurrence prediction from weakly annotated H&E stained tissue images. These images (CPCTR data) were collected from patients undergoing radical prostatectomy. I primarily worked on implementing attention based multiple instance learning technique in pytorch for predicting class probabilities based on features extracted by the VGG16 pretrained network. Attention scores were also generated to mark the responsible tissue regions for interpretation. The repository is available on my github.
Prior to that in the summer of 2018, I worked on image processing with Prof. Mrigank Sharad in his start-up AgNEXT. I implemented a grain (rice/wheat) segmentation algorithm in opencv and C++ based on polygonal approximation of contours and their gradients. Its purpose was to isolate clusters of (touching) grains in a photograph. This computational method was used as a preprocessing step to get the size and color statistics of a grain sample.
Prior to that in the summer of 2018, I worked on image processing with Prof. Mrigank Sharad in his start-up AgNEXT. I implemented a grain (rice/wheat) segmentation algorithm in opencv and C++ based on polygonal approximation of contours and their gradients. Its purpose was to isolate clusters of (touching) grains in a photograph. This computational method was used as a preprocessing step to get the size and color statistics of a grain sample.
