Home Research Projects CV Contact

Puranjay Datta

Hi! I'm a fifth-year undergraduate student pursuing a dual degree program in Electrical Engineering (M.Tech + B.Tech) at IIT Bombay, with a minor in Computer Science. My academic and research interests lie in the fields of AI and intelligent agents, stochastic control, online learning, and game theory.

Since the summer of 2023, I have been working with Professor Sharayu Moharir on research related to inference using hedge methods. Additionally, I had the opportunity to be a MITACS Globalink Research Intern at the University of Calgary under the supervision of Hatem Abou Zeid. During this internship, I delved into the domain of transfer learning using IsaacGym, gaining valuable experience in this area.

Furthermore, I am currently collaborating with Professor Shivaram Kalyanakrishnan to enhance and optimize his NeurIPS Reconnaissance Blind Chess Agent bot, with the aim of preparing it for future competitions. I also had the privilege of working alongside Professor Swaprava Nath on the Envy-Free Cake Cutting problem as part of an advanced game theory course, an experience that I found immensely enjoyable and intellectually stimulating.

Beyond my academic pursuits, I had the privilege of serving as a Teaching Assistant for a national-level NPTEL Digital Signal Processing course, where I worked alongside Professor Vikram Gadre to ensure the smooth delivery of the course. In my leisure time, I am an avid sports enthusiast and follow a wide range of sports, including cricket, chess, Formula 1, table tennis, and basketball. I also enjoy actively participating in some of these sports during my spare time.

profile photo

Drafts in preparation




Get in touch !



I am always happy to chat and assist you with any questions or concerns you may have. Feel free to reach out to me and I will get back to you as soon as possible.



Research Experience


Predistortion in Power Amplifier

Experimented with Iterative Learning Control, Vector Switched Models on industrial amplifiers for various bandwidths and compared based on ACLR, SNR metrics.


Transfer learning in NVIDIA isaacgym

Analyzed the Adversarial Motion Priors in training humanoid agents using motion datasets, investigating transfer learning to another task using one-shot transfer learning


Neurips Reconnaissance Blind Chess Agent
Presentation / Report

Replay buffer development, opponent modeling, and a scoring system adaptation were employed to assess and enhance the Fianchetto bot's(modelled as a POMDP) performance in chess, focusing on threat detection and strategic move selection.


Networked Fairness in Cake Cutting
Presentation / Report

Pioneered a novel approach to traditional cake-cutting with networked agents for efficient envy-free allocations using moving knife procedures Austin Cut and Brams Taylor Zwicker procedure


Inference using Online Learning
Presentation / Report

Hierarchichal inference using modified hedge with a non convex lipschitz loss in a 2D continuous experts setting utilizing insights from Online Recursive Weighting algorithm.

Projects


Multiarm Bandits
Presentation 1 / Presentation 2

Involved multi-armed bandit best arm identification, implementing successive elimination and median elimination algorithms, adapting confidence bounds for grouped bandits with constraints, and analyzing the stopping time complexity of a D tracking variant against hardness measures with a focus on helpful arm characterization.


Generative model for Weak Supervision
Report

leveraged the Snorkel framework to train a generative model for learning relationships among multiple labeling functions, which, when combined with discriminative models like logistic regression and recurrent neural networks, significantly enhanced accuracy in tasks like salary prediction and Twitter sentiment analysis.


Wavelets in Convolutional Neural Network
Presentation / Report

Developed a novel sparse neural network combining LSTM and wavelet decomposition for predicting atmospheric profile by implementing Level-2 decomposition along with LSTM and separate k-band, v-band training to improve the accuracy.


Shadow Removal and Detection
Report

Implemented the water-filling diffusion algorithm and k-means to equalize the global and local shadow background and compared with a Stacked CGAN on an Image shadow triplets (ISTD) dataset with annotations in the form of the shadow mask.


Temperature Control using Pulse Width Modulator
Report

Fabricated a functional prototype with temperature control, integrating a PCB layout in Eagle and creating a dynamic temperature display using an Arduino-based NTC thermistor with feedback-controlled heating.


Optical Flow Computer Vision
Presentation / Report

Tested Lucas Kanade Optical flow on various test cases showcasing its strength and weakness

page design source: Jon Barron , Image by user3802032 on Freepik