research
papers
* in alphabetical order
- Anoop Kunchukuttan, Maulik Shah, Pradyot Prakash, Pushpak Bhattacharyya, “Utilizing Lexical Similarity between Related, Low-resource Languages for Pivot-based SMT.” IJCNLP, 2017
- * Stanley Bak, Sergiy Bogomolov, Thomas A. Henzinger, Taylor T. Johnson, Pradyot Prakash, “Scalable Static Hybridization Methods for Analysis of Nonlinear Systems.” HSCC 2016
projects
Breaking Robust Adversarial Classification (retrieve here)
Spring 2018
Under Prof. Dimitris Papailiopoulos, UW-Madison
- Robust Manifold Defense is state-of-the-art adversarial classification algorithm which works by projecting on the space of GANs
- Developed the first algorithm to break the classifier & brought down the classification accuracy by 35% (more work in progress)
Fuzzy Iterative Machine Teaching (retrieve here)
Spring 2018
Under Prof. Jerry Zhu, UW-Madison
- Studied the inverse problem of ML–Machine Teaching–where the aim is to learn a target parameter vector in minimum steps
- Derived robust bounds for the minimum steps needed to converge under noisy and missing data settings for different losses
Autoencoders & Generative Adversarial Modeling
Spring 2018
Under Prof. Rebecca Willett, UW-Madison
- Implemented autoencoders and generative adversarial nets for MRI image denoising
- Worked on a unique approach to invert a neural network using Neumann series in operator space
GPU profiling of Deep learning frameworks (retrieve here)
Fall 2017
Under Prof. Aditya Akella, UW-Madison
- Analyzed deep learning libraries using their intermediate representations (with ONNX) and profiled their GPU performances
Effect of Segmentation and Encoding on Machine Translation
2016 - 2017
Under Prof. Pushpak Bhattacharyya, IIT Bombay
- Improved transliteration and translation with use of pivot-based modeling, byte-pair encoding and orthographic syllabification
- Paper published at IJCNLP 2017 (retrieve here)
Statistical Shape Analysis of Images (retrieve here and here)
2016 - 2017
Under Prof. Suyash Awate, IIT Bombay
- Created a similarity measure between images and their contours alongside a novel objective function
- Used that to segment MRI images (with a shape prior) by performing Riemannian PCA on high dimensional unit spheres