Hi, I'm Steven Chen.

I'm a computer science student at the University of Toronto, doing machine learning research with Project Neura and UTMIST. Most of my work is about reliability: getting models to be honest about what they do not know, and finding out where training methods break before anyone builds on top of them.

In medical imaging I co-first-authored MIPCandy with Tianhao Fu, a modular PyTorch framework for medical image processing, and I work on SegWithU, which treats uncertainty as perturbation energy so a frozen segmentation model can flag its own failures in a single forward pass.

Before that I spent two years on Forward-Forward, Hinton's proposal for training networks without backpropagation. I built the strongest FF network I could and then used it to show the method does not scale: on real images an architecture-matched backprop baseline stays ahead and the gap widens with task complexity, and the synthetic benchmarks the field uses to argue otherwise have the sign backwards. The paper is on arXiv, and I wrote up what I found.

On the markets side, most of my work now goes into IV Surface Explorer, a dashboard I built because I wanted one: implied volatility surfaces, term structure, skew, gamma exposure, open interest walls. Earlier there was ATLAS, which borrows computer vision methods for stock pattern recognition, and Portfolio Manager, built for the Wharton WGHS competition.

I run Amplimit, the research organization that carries the affiliation on my papers, and I build Ampdraft, a writing tool that gives documents real version control. I also work part-time as an engineer at Kabuda, building AI and web products.