Machine Learning
LLM as a Multi Agent Operator
Self-Supervised and Weakly Supervised Representaton Learning
Existing methodologies of SimCLR, beta-VAE, VICReg were replicated and applied to new datasets to test generalization. This study explores disentanglement within the context of representation learning in new domains.
This project was done as an independent study under the guidance of Dr. Richard Lange at RIT’s Computer Science Department.
Cross-Organizational Continual Learning of Cyber Threat Models
An end-to-end continual learning system was developed using experience replay that learned from an online data stream of network traffic. This project enabled continual learning from multiple sources of data over time using data homogenization and further employed uncertainty sampling techniques to reduce the amount of labeled data required.
This project was done under the guidance of Dr. Shanchieh Jay Yang at the RIT ESL Global Cybersecurity Institute.
Project poster and Github repository.
Meta Learning of Emotions from EEG
Model Agnostic Meta Learning was implemented for EEG emotion recognition capable of adapting quickly to new subjects with minimal data. The machine learning models use few-shot learning with normalization and alignment techniques to handle EEG variability across subjects.
This project was done in collaboration with the RIT Neurotechnology Exploration Team.
High Power Microwave Parameter Sensitivity Analysis Using Advanced Machine Learning Techniques
Active learning surrogate models were designed that could optimize high power microwave device parameters, reducing computational cost and time needed to simulate and design each device. An interactive query system was introduced for obtaining human labels for model predictions that had high uncertainty.
This project was done under the supervision of Dr. Ashar Ali at the Air Force Research Laboratory Scholars Program.