Paul Janson

PhD Student

About Me

I am a PhD student at Mila Quebec AI Institute and Concordia University in Canada with a passion for deep learning and computer vision research. My primary research interest lies in improving the efficiency of training foundation models by leveraging prior knowledge through continual learning, meta-learning, and model reuse. I am being supervised by Dr.Eugene Belilovsky. Before coming to Mila and Concordia , I worked as a research intern at KAUST in Vision CAIR group under the supervision of Dr.Mohammed Elhoseiny. I did my undergraduate at University of Moratuwa, Sri Lanka with thesis in reconstructing 3D objects with few views with Dr. Thanuja Ambegoda.

Download my CV .

Interests

Continual learning
Meta learning
Foundation models
Computer vision

Education

PhD in Computer Science, 2024-2028
Concordia University
Masters in Computer Science, 2023-2025
Concordia University
BSc in Computer science and Engineering, 2018-2023
University of Moratuwa

News

2026

[2026-06-01] Started a Research Internship at Microsoft Research, Cambridge, UK
[2026-06-01] 1 paper accepted at the ICML 2026 2nd Workshop on Connecting Low-rank Representations in AI
[2026-05-01] 1 paper accepted at ICML 2026 Main Conference
[2026-05-01] PyLO accepted at MLSys 2026

2025

[2025-07-09] We released PyLO an open source optimization library to enable usage of learned optimizers in PyTorch
[2025-07-08] 2 papers accepted at ICML 2025 Workshops
[2025-05-12] 1 paper accepted at CoLLAs 2025 Main Conference (oral)

2024

[2024-09-29] 1 paper accepted at ECCV 2024 Workshop
[2024-06-01] 1 paper accepted at CVPR 2024 Main Conference
[2024-05-01] Fast tracked to PhD in Computer science from MS under Dr.Eugene Belilovsky

2023

[2023-10-03] 2 papers accepted at ICCV 2023 Main conference and workshop
[2023-09-01] Started my MS in Computer Science at Mila and Concordia University under Dr.Eugene Belilovsky

2022

[2022-10-20] 1 paper accepted at NeurIPS 2022 Workshop

Publications

Learned Subspace Compression for Communication-Efficient Pipeline Parallelism

In ICML 2026 2nd Workshop on Connecting Low-rank Representations in AI.

Paul Janson, Edouard Oyallon, Eugene Belilovsky

MAPL treats inter-stage activation compression in pipeline parallelism as a learnable orthogonal projection under Stiefel manifold constraints, letting each stage adapt its own task-optimal subspace. Combined with factorized anchor embeddings and residual vector quantization, it achieves high compression with negligible performance loss across LLaMA models from 150M to 1B parameters.

Learned Subspace Compression for Communication-Efficient Pipeline Parallelism

PyLO: Towards Accessible Learned Optimizers in PyTorch

In Proceedings of Machine Learning and Systems (MLSys) 2026.

Paul Janson, Benjamin Therien, Quentin Anthony, Xialong Huang, Abhinav Moudgil, Eugene Belilovsky

A PyTorch library that makes learned optimizers accessible to the broader ML community with CUDA-accelerated implementations with substantial speedups (5x improvement for ViT training) and integrates seamlessly with existing PyTorch workflows, enabling practical application of learned optimization to real-world large-scale tasks.

PyLO: Towards Accessible Learned Optimizers in PyTorch

Stabilizing Native Low-Rank LLM Pretraining

In International Conference on Machine Learning (ICML) 2026.

Paul Janson, Edouard Oyallon, Eugene Belilovsky

We identify the uncontrolled growth of weight-update spectral norms as the key instability in training natively low-rank LLMs, and introduce Spectron, a lightweight spectral renormalization technique that enables stable end-to-end factorized pretraining with compute-optimal scaling laws and improved inference efficiency.