Multilingual and Multimodal NLP

My experiences with multimodal multilingual language models came from my research with Professor Alex Kelly at Carleton University. Our hypothesis on the usefulness of multimodality in multilingual learning is based on multimodal learning in humans: Since humans learn languages using visual and textual information, models trained with visual grounding should also behave more human-like than those without. As our previous work dealt exclusively with sequence models, particularly the LSTM, I hope to examine this hypothesis on attention-based architectures.

Interpretability

My experiences with mixed integer linear programming (MILP) came from my research with Professor Thiago Serra at Bucknell University. For our research, I developed several MILP programs using constraints to simulate the inner-working of the perceptrons, through which I was able to detect their stably active and stably inactive neurons as well as counting their number of linear regions. I hope to extend this method to current language models to better understand how they behave.

BERTology

I have been tinkering with BERT and models deriving from BERT, particularly focusing on the summarization ability of BERT’s attention layers. I am also keen on exploring new BERT-based architectures, and much of my Kaggle projects have been about me trying out transformer models on HuggingFace (RoBERTa, BERT, BigBird, Longformers, etc).