Hi! I’m Cornelius, a first-year PhD student in Electrical and Computer Engineering at McGill University and Mila, working with Tal Arbel. My research focuses on generative modeling for medical imaging — currently 3D diffusion models for brain MRI. My broader interests include diffusion and generative models, representation learning, computer vision, and deep learning.

Community & Teaching

Beyond research, I’m an active member of the UNIQUE Student Affairs Committee, where I help organize student conferences and social events like the UNIQUE Student Symposium (USS) and semesterly get-togethers featuring research presentations and networking opportunities for the NeuroAI community. I also love teaching and mentoring. Being a TA has been one of my favorite roles during my studies because it allows me to share knowledge, guide students, and help others develop their own understanding of deep learning. Whether it’s through formal teaching or casual discussions, I find a lot of fulfillment in helping others grow.

What Keeps Me Inspired

When I’m not doing research, you’ll likely find me doing outdoors or being active. I love running, rollerblading, bouldering, ice skating (during winter), biking, or simply being in nature. I’m currently training for my first marathon, which I once thought would be impossible. Rollerblading, especially urban and wizard skating, is one of my favorite hobbies. It’s so much fun experimenting with gear and different skate setups too. I’m part of Montreal’s Rolling Tribes community, which has been a great way to stay active and meet people.

I also enjoy reading fantasy and sci-fi, listening to live music, and thinking about urban design. I care about cities being built around people instead of cars, which in my opinion drastically enhances daily life. Fun fact: I can solve a Rubik’s Cube in under 30 seconds!

Projects

Generative Modeling for Medical Imaging

Developing 3D diffusion-based generative models for brain MRI. More details to come as the work is published.

Visual Enhancement of Whole Brain Slide Images from Z-Scanning Microscopes with Deep Style Learning

In this master’s thesis project, we developed the Z-Stack Enhancement (ZSE) pipeline, a deep learning framework implemented using Pytorch Lightning designed to enhance whole-brain microscopy images. This approach transfers the visual style from an in-focus reference image to out-of-focus images using a U-Net architecture combined with Adaptive Instance Normalization (AdaIN). The method effectively deblurs microscopic images, enhancing clarity and making it easier to study detailed brain structures.

Key contributions:

• Development of a novel deep style learning framework inspired by neural style transfer, specifically adapted for microscopy imaging.
• Introduction of two novel 3D loss functions optimized for image stacks, resulting in more realistic reconstructions of brain tissue.
• Evaluation on two datasets, including super-high-resolution brain z-stacks from Research Center Jülich and microscopy images of Leishmania parasites, achieving state-of-the-art results with fewer parameters and improved generalization compared to existing methods.
• Public release of trained models and PyTorch implementation on a GitHub Repository.

Playing Go with Recurrent Neural Networks

In this bachelor’s thesis, we implemented neural network models using TensorFlow for playing Go via imitation learning, training on extensive datasets of expert human games. The project’s goal was to evaluate different neural network architectures regarding their learning efficiency and effectiveness in mastering the game.

Key contributions:

• Implementation and comparison of two architectures:
  • A RNN model consisting of a CNN backbone with an integrated ConvGRU layer for context summarization from past board states.
  • A CNN-only approach, inspired by AlphaGo Zero, utilizing additional input channels to encode past states without recurrence.
• Demonstration that the RNN-based model learned more efficiently compared to the CNN-only model, though both ultimately achieved comparable performance.
• Providing insights into the significance of temporal modeling for decision-making processes in reinforcement learning contexts.