This projects tackled major challenges in wearable technology: the intensive labeling of biosensor data and the subjective nature of stress measurements. This study demonstrates that Self-supervised Pretrained models outperform Purely Supervised Trained models while utilizing less than 5% of the annotations while using multi-modal biosignals, and 30% of the annotations while using single-modal biosignal.

This project examines the impact of racial bias in self-supervised learning models due to unbalanced pre-training datasets. By simulating diverse racial compositions, the study assesses bias in medical imaging datasets used for predicting breast cancer stages via MRI, aiming to enhance diagnostic fairness.

Focused on developing a custom loss function that incorporates fairness regularization with Demographic Parity, this initiative seeks to balance model predictions across demographic groups, promoting equity in healthcare outcomes.

This project centered on creating a personalized food recommendation system for individuals with severe communication impairments, such as locked-in syndrome. Utilizing electroencephalography (EEG) signals, machine learning models were developed to predict individual affective responses to different foods. The project incorporated the TOPSIS multi-criteria decision analysis method to tailor meal planning based on affective reactions, nutritional content, and energy needs, thereby advancing expertise in affective computing and machine learning.

This project focused on improving headline generation for the low-resource Bengali language by enhancing encoder-decoder models with auxiliary information such as image captions, categories, and topic words. This approach significantly outperformed baseline models in both few-shot and regular settings, effectively addressing the fixed length limitations of pre-trained models like BERT