Geometry-aware quantization for the Muon optimizer
Muon's updates are orthogonal. We quantize them to a few bits without breaking that geometry, cutting the memory and communication cost of training large models.

Machine learning that is efficient enough to scale and rigorous enough to trust.
I'm a CS Master's student at Cornell University, where I work with Prof. Christopher De Sa on low-precision, geometry-aware optimization for training large models. Before Cornell I spent two and a half years at the Vector Institute building representation-learning and causal-inference methods with oncologists at UHN, and a year and a half at Goldman Sachs engineering production systems.
Fewer bits per update when training large models; fewer unchecked assumptions when the question is causal.
Muon's updates are orthogonal. We quantize them to a few bits without breaking that geometry, cutting the memory and communication cost of training large models.
Wearable and survey time series, a latent-confounder-aware tsFCI, and a synthetic benchmark to prove it recovers the right graph.
Foundation-model embeddings and vision transformers on whole-slide pathology across 4 cohorts, to predict survival and estimate treatment effects.
Research at the Vector Institute and Cornell University, engineering at Goldman Sachs. The same discipline each time: make the hard problem measurable, then build it so it lasts.
Developing few-bit quantization for Muon optimizer updates, with group-theoretic codebooks that preserve orthogonality, to cut the memory and communication cost of LLM training.
Built vision-transformer and causal-inference pipelines on whole-slide pathology across 4 patient cohorts with UHN oncologists, to predict survival and estimate treatment effects.
Designed and launched an automated system processing 1M+ transactions a day, and re-engineered legacy tax systems for 10× runtime gains.
Graduate study in large-scale machine learning and reinforcement learning, advised by Prof. Christopher De Sa.
Ranked top 1 in the graduating class; named top student in AI.
Public code, from a diffusion-model paper to a full-stack marketplace. Everything is on GitHub.
Teach a latent diffusion model a new subject from a few photos, with caption-initialized embeddings instead of a cold start.
An autoencoder that predicts student answers, with item-response-theory priors in its latent space.
A short-term rental platform for homeowners and tenants: accounts, listings and a full reservation lifecycle, built as a single-page app over a REST API.
Four years of lecture notes typeset in LaTeX and published for classmates, from first-year calculus to deep learning. Still the notes people ask me for.
Mostly for the joy of it. The Royal Conservatory's ARCT diploma (with first class honours) came along the way, and for five years I hosted monthly concerts at a seniors' residence in Richmond, BC.
At eight hours a day, it would take nearly a month to get through.
Open the playlistsBadminton and tennis whenever there's a free court, and the odd round of golf. Swimming goes back to high school, where I also earned my National Lifeguard certification. Winters mean skiing, and Whistler Blackcomb is still my favourite mountain. Weekends slow down: a board game or a glass of wine.
Longer vacations are for travelling, and I tend to plan each trip around a city's museums. One museum at a time, the map below keeps filling in, across China, the U.S. and well beyond.
Research, hard engineering problems, or a music recommendation. My inbox is open.
wei at cs.toronto.edu