I am a pre-final year undergraduate student pursuing a major in Computer Science with a Minor in Data Science at
My research interests include Vision, Autonomous Systems, Reinforcement Learning and Causal ML. I am also interested in High Performance Computing and Parallel File Systems.
Research Intern | North Eastern Space Applications Centre
May'23 - August '23
The project aimed to develop predictive models for rainfall forecasting, improving upon the prediction horizon of existing weather models. A data pipeline was created for pre-processing geospatial datasets and integrated with a UNet-based architecture. Experiments were conducted combining genetic algorithms and gradient-descent based optimizers to improve accuracy.
Student Researcher | APPCAIR
June '23 - Present
Designing a non-chaotic neural network pruning strategy that also preserves feature importances. Investigating Granger Causality and comparing against L0, L1, Rank-based and other state-of-the-art algorithms.
Student Researcher | Data, Systems and High Performance Computing Lab
April '23 - Present
Part of a team that is maintaining a 32-node cluster consisting of rack servers, desktop workstations and Raspberry Pi's. Designing an algorithm for optimal file-volume mapping in BeeGFS, a parallel file system. Incorporating adaptive striping, access pattern analysis and file-size awareness.
Does Varying BeeGFS Configuration Affect the I/O Performance of HPC Workloads?
Arnav Borkar, Joel Tony, Hari Vamsi K. N, Tushar Barman, Yash Bhisikar, Sreenath T. M. and Arnab K. Paul
Developing a mars rover as part of the University Rover Challenge. We're responsible for the rover's autonomous navigation.
Worked on a real-time object detector for the rover using YOLOv3 trained on our custom dataset and ran it on a Jetson Nano. Achieved integration with ROS using darknet_ROS. Designed a PID-inspired mechanism to navigate directions based on arrow directions captured in a monocular camera feed. Integrated GPS-based navigation mechanism with the tech stack to traverse between local GPS coordinates.
My implementation of the Decision Transformer. I modified the architecture to use an LSTM to compare training accuracies and performance in a non-Markovian setting. Used a sequence of Reward, State and Action tokens to condition the model and predict the next expected reward and optimal action. Validated results against model-free offline RL Baselines in Mujoco's Hopper environment.
Segmentation on Arbitary Image-Text prompts using Clip
Used OpenAI's Clip to generate embeddings of the input image/text prompt and performed Featurewise-Linear Modulation. Inspired from the UNet, trained a light-weight transformer-based decoder on top of the embeddings and integrated encoder activations to generate the binary segmentation mask in a supervised setting.
Data Prefetching for Edge Deep Learning Workloads
Semester Project for the Course - Data Storage Technologies and Networks
Desigining a comprehensive framework to optimize convergence time for Deep Learning training workloads on edge devices. We're implementing a prefetching mechanism accounting for spatial and temporal locality to reduce I/O overheads and overall training time. Apache Kafka is being used to stream data from a distributed file system to the compute nodes.
This template is a modification to Jon Barron's website
and a fork of Hardik Shah's website.
Find the source code to my website here.