I like rice pudding

About Me

I am a PhD student at Mila advised by Prof. Sarath Chandar.
My research broadly focuses on understanding the principles behind how deep learning models work and uncovering their underlying training dynamics. My current interests lie in sequence modeling and memory augmented neural networks, in particular the use of structured memory and how to leverage it for adapting to distributional shifts or enabling more robust information flow. I furthermore investigate how these can lead to better transfer and enable reliable retrieval of information within models. On the side, I also dabble in more applied research in natural language processing.
My research is (or has been) generously supported through funds from NSERC, FRQNT, Hydro-Québec, the 日本学術振興会 (Japan Society for the Promotion of Science) among other sources. I thank them for their broad support which have enabled me to conduct research into topics I personally deem valuable.

Investigating the Effects of Architectural Inductive Biases on Hallucination [Link]
ZETA: Leveraging Z-order Curves for Efficient Top-K Attention [Link]
How Well Can a Long Sequence Model Model Long Sequences? [Link]
Do Large Language Models Know How Much They Know? [Link]
Context-Aware Assistant Selection for Improved Inference Acceleration with Large Language Models [Link]
Predicting the Impact of Model Expansion through the Minima Manifold [Link]
Towards Practical Tool Usage for Continually Learning Large Language Models [Link]
EpiK-Eval: Evaluation for Language Models as Epistemic Models [Link]

I am actively seeking motivated students for collaboration. If you are passionate about research and eager to work on innovative projects in deep learning, consider reaching out. In addition, please fill out this form - I will follow up if there is a strong fit and ample resources. In the event that you satisfy one of the following criteria, please reach out directly (by e-mail) as there may be additional ways to increase the chances of collaboration:

You satisfy the eligibility requirements for a NSERC USRA.
You study at one of Mila's partner unversities and must conduct research to satisfy degree requirements.
You have external funding to support an internship/research stay and are not a resident of Canada.

As a general (but neither necessary nor sufficient) overview of what I personally look for in terms of collaborators

Familiarity with standard deep learning libraries: You should able to train and evaluate some standard networks and models on a task related to your topic of interest. Additionally, you should be able to read and explain code written by others.
Effective communication: Being able to compose clear and coherent ideas both in verbal and writtern form is imperative for disseminating research properly. An easy way to demonstrating this is providing a short research proposal or report.
Diligence and persistence: A somewhat more personal preference, I like observing when an individual can incorporate suggestions into their research ideas as well as counter those that are likely unnecessary.

Nevertheless, everyone has different experiences and I encourage those from a non-conventional background (specifically non-engineering or math based) to contact me if you are interested but might not meet all of the above criteria.

Please use my e-mail for contact. I check it daily and respond in a timely fashion (between 1 to 2 working days depending on the urgency).

Outside of research, I read books and cook (particularly seafood). I faithfully abide by the principle of 腹八分目 (hachi hachi bun me - eat until you are eight parts full). I also play badminton, swim, fish and snowboard (in the winter).