Continual Optimistic Initialization for Value-Based Reinforcement Learning (COIN)
NEW

Project currently under paper construction phase. Links and updates will be added in the coming months

• Designed COIN exploration strategy overcoming optimistic initialization limitations in non-stationary environments.• Showcased higher average returns over existing strategies across five out of 6 benchmark domains in OpenAI Gym, illustrating the superiority of COIN strategy.• Conducted ablation studies to analyze individual hyperparameters.

Hybrid Software Architecture for AI Rescue robot
NEW

Colaboration between Texas A&M University and Stanford University

• Designed a ROS architecture for Survivor Buddy, a pioneering collaboration of TAMU & Stanford University.• Leveraged observer design pattern to automate 5 human robot interactions, spanning 2K+ lines of code.• Devised multi-threaded algorithm for parallel image & audio processing, reducing latency by 95%(≈ 0.1sec).

Presentation GitHub Demo Website

Improving Generalization of Deep AUC Maximization for Medical Image Classification

Dr. Tianbao Yang, Texas A&M University

• Improved the generalization ability of Deep AUC Maximization for medical image classification.• Performed ablation studies to investigate the impacts of augmentation, balancing imbalanced datasets, addressing distributional shift, and tuning hyperparameters.

PAPER GitHub

SkyTrace: Aerial Artistry Tool

Sketch Recognition Lab, Texas A&M University

• Implemented fingertip detection and locus tracking for crafting air-drawn sketches from video data• Used OpenCV’s Mediapipe for precise hand-tracking & developed complex multi-stroke air sketch algorithms• Pioneered a concept of a seamless scene delimiter, crucial for executing multiple strokes cohesively

GitHub

PayFast: Online Wallet

Designed an application for TAMU's internal payments portal

• Developed a Ruby-based online wallet application to streamline transactions within Texas A&M’s portal• Integrated multi-currency, detailed history, and seamless card management for enhanced user convenience

GitHub Demo Twitter

A voting approach to uncover multiple influential spreaders on weighted networks

Dr. B.S Panda, IIT-Delhi

• Optimized WVoteRank algorithm for weighted graphs to identify influential nodes• Improved the scoring metric and loss function by additionally incorporating second neighbours• Presented findings to 100+ research scholars in a seminar conducted by Mathematics department at IIT-Delhi

Presentation