Publications
2020
Lower Bounds for Policy Iteration on Multi-action MDPs
Kumar Ashutosh*,
Sarthak Consul*,
Bhishma Dedhia*,
Parthasarathi Khirwadkar*,
Sahil Shah*,
Shivaram Kalyanakrishnan*
IEEE Conference on Decision and Control 2020 *Equal Contributions
Paper
Devised a family of n-state, k-action MDPs to obtain a strong lower bound of $\Omega(k^n/2)$ for policy iteration. Furthermore, we generalised the existing constructions of 2-action MDPs to k-action MDPs to scale lower bounds by a factor of $k$ for some common deterministic variants of policy iteration, and by $\log(k)$ for the corresponding randomised variants.
Compressed Sensing Approach to Group-testing for COVID-19 Detection
Sabyasachi Ghosh,
Rishi Agarwal,
Mohammad Ali Rehan,
Shreya Pathak,
Pratyush Agrawal,
Yash Gupta,
Sarthak Consul,
Nimay Gupta,
Ritika Goyal,
Prof. Manoj Gopalakrishnan and
Prof. Ajit Rajwade
IEEE Open Journal of Signal Processing 2021
Paper | Code | 
Official Website
We propose a single-round pooled testing approach, called Tapestry, with an application to detect SARS-CoV-2 viral loads using quantitative RT-PCR that shortens testing time and conserves reagents and testing kits. A combination of combinatoric approaches (such as COMP) and compressed sensing algorithms (such as Sparse Bayesian Learning) with a novel noise model for PCR allows us to be able to recover both the status and estimated viral load of the samples. An accompanying Android application has been developed for easy implementation at testing centres.
Tapestry has been approved by the Drugs Controller General of India (DCGI) for commericial deployment after extensive testing!
Research Projects
2020
Style Transfer on Unpaired Music Clips
Bachelor’s thesis guided by Prof. Subhasis Chaudhuri
Style transfer was introduced for images in 2015 and has been quickly commercialized in the entertainment industry. Style transfer for audio has not enjoyed the same success. A cyclical Wasserstein GAN was trained to tranform the vocals of songs to mimic the singer of an unpaired target source. Vocals were separated from YouTube music clips using a ConvTasNet trained on the DSD100 dataset.
2019
Analysis of Lower Bounds on Simple Policy Iteration for K-action MDPs
Guided by Prof. Shivaram Kalyanakrishnan
Project Paper
Simple Policy iteration (SPI) is a type of policy iteration where the strategy is to change the policy of exactly one improvable state to an arbitrary improving action at every step. Melekopoglou and Condon [1990] showed an exponential lower bound on the number of iterations taken by SPI for a family of 2-action MDPs. We generalized the result to obtain a lower bound for the family of k-action MDPs.
Leveraging Reinforcement Learning for Semantic Segmentation
Guided by Prof. Amit Sethi
Developed an hierarchical segmentation agent trained using the REINFORCE policy gradient algorithm to achieve an mIOU of 53.62% on the PASCAL VOC 2012 dataset.
Segmentation of Lacunar Objects from Ultra High Resolution µCT Bone Scans
Guided by Prof. Ralph Müller, Internship Project at ETH Zürich
Lacunae are small cavities within the bone matrix, which each contain an osteocyte. Osteocyte mechanics are governed by lacunar mechanics by phenomena such as mechanosensation,
where the shape of the lacuna is important. In an effort to understand difference in osteocyte mechanics in healthy
and patients suffering from rare forms of osteoporosis (specifically pregnancy-associated osteoporosis), the detection
and characterization of these lacunae can serve as a biomarker. The classical approach involves simply thresholding the scan
on a single value of bone mineral density. This method fails to capture many small lacunae, which is believed to be the main
area of interest in explaining the diseases of interest. We propose a 2 stage approach of adaptive thresholding followed by a classifier to remove noise structures, to obtain good estimates of the lacunar location and shape.
2018
Semantic Segmentation of Medical Images
Guided by Prof. Amit Sethi
Project Page
Segmentation is a per-pixel clsssification task that is more complicated than classification. We explore the task of semantic segmentation of medical images. This has immense applications in diagonosis and pathology. The work was used in establishing the baseline in the MoNuSeg Challenge at MICCAI 2018.