Research & Projects

Project 1

Hierarchical Optimization of Neurostimulation for Spinal Cord Injury

This project looks at the optimization of neurostimulation parameters for joint muscle activations following a traumatic spinal cord injury in rats. This work is the main aim of my master's thesis and has been conducted under the joint supervision of Professors Marco Bonizzato and Guillaume Lajoie at Polytechnique Montreal and Mila - Quebec AI Insitute.

Objectives

Our goal was to find a way to incorporate multiple muscles into the stimulation procedure, a problem due to the increasing complexity of adding more parameters. This is important as recovery of movement is not sufficient when only one muscle is recovered following paralysis, prompting the need to learn how to recover function from multiple movements simultaneously.

Results

We were able to design a hierarchical learning algorithm based on Gaussian Processes that was able to disentangle information. This means that we were able to, in essence, train multiple different models simultaneously with only one response signal. This structure could then be applied to neurostimulation settings with multiple optimization points as well as any hierarchical learning setting.

Project 2

Hallucination Detox for Large Language Models

Hallucinations are a problem that plagues the world of Large Language Models such as ChatGPT and Meta's Llama models. This issue is characterized by the production of factually inaccurate or irrelevant outputs, which could be harmful depending on the topic the LLM is talking about. In collaboration with fellow Mila colleague Shahrad Mohammadzadeh and under the supervision of Professors Marco Bonizzato, Reihaneh Rabbany, and Golnoosh Farnadi, we aimed to shine light on this kink in the world of LLMs

Objectives

Our target was to first understand why hallucinations may have been happening, something that other LLM researchers have failed to do, and see if it could be fixed within the training protocol. This differed from most research in the field as the status quo is in the direction of tackling hallucinations after model deployment as an additional algorithm. We were the first to ask ourselves and conduct research on development of this phenomenon throughout training.

Results

We found that current LLM models are trained to optimize their loss function, but this does not correspond to optimizing factual performance. Observing the training dynamics of Pythia, we observed that hallucinatory performance was not being affected by the training protocols and would fluctuate near randomly as training proceded. In response to this, we created a novel training protocol - Sensitivity Dropout (SenD) - aimed at reducing the presence of these oscillations and improving end-model hallucination performance. Additionally, we developed a new hallucination detection method - Efficient EigenScore (EES) - which approximates a state of the art unsupervised detection technique in quadratic time rather than cubic. This work was recognized by the community as a substantial contribution and was published at the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025).

Project 3

Unveiling Lateralization of the Brain

The brain is inarguably the most complex and lateralized system in animals, meaning that it is the most different when split down the middle compared to other organs. Among animals, humans have the most lateralized brain of all with significant evolutionary cost accrued to attain this type of brain. Within us, we find the temporoparietal junction, a region of the brain which is argued to be one of the most laterlized regions within the human brain and is thought to be a key factor in human defining abilities. For these reasons, Professor Danilo Bzdok at McGill University and Mila - Quebec AI Institute and myself alongside international colleagues tasked ourselves with understanding this region.

Objectives

With a unique causal transcranial magnetic stimulation strategy, we aim to look at what human defining capabilities are located within the temporoparietal junction and how these differ between the left and right side of the brain.

Results

We found that the temporoparietal junction is a highly lateralized region that, contrary to popular belief within the scientific community, is not differently involved in social tasks between the right hemisphere compared to the left. However, another uniquely human task - language - was found to be significantly different between these two hemispheres in the temporoparietal junction.

Page updated

Report abuse