Research hosted at IBM Research in the ARMADA doctoral network, co-funded by Switzerland’s SERI and the European Union.

Developed Hate-CLIPper, a CLIP-based hateful-meme classifier that models image–text interactions with a feature interaction matrix, reaching 85.8 AUROC on the Hateful Memes Challenge and surpassing the human baseline of 82.65.

Developed MuCoT, a multilingual contrastive training method for low-resource QA that augments examples via translation and a contrastive loss, improving mBERT on the Google ChAII benchmark.

Thesis: Towards Learning Efficient Multilingual and Multimodal Representation.

Relevant Coursework: 

• ML701 - Machine Learning (ACL Workshop Paper - Multilingual Question Answering Project)

• CV701 - Computer Vision

• CV702 - Geometry for Computer Vision

• CV703 - Visual Object Detection and Recognition (Object Detection with Vision Transformers Project)

• ML708 - Trustworthy AI (EMNLP Workshop Paper - Mulitmodal Hateful Meme Classification Project)

Relevant Coursework: 

• Object Oriented Programming

• Data Structures and Algorithms

• Web Technologies

• Data Base Management System

• Software Engineering

• Big Data

• Data Mining