Steven Luo

I'm a 4th year UofT CS undergrad.

Since my first endeavour into AI in 2018, I've been advised by some wonderful people -- Jonathan Binas and Pegah Kamousi at Vivid machines; David Lindell at UofT DGP; HKU EEE's Ngai Wong; Matthias Niemeier at UTSC CoNSens.

Email / Resume / Scholar / Twitter / Github

Research

My interests stem from the root vision of replicating human intelligence both as:

An end - to understand what makes human intelligence so unique, if unique at all.
A means - to scale human productivity for the most difficult and important problems in the world.

Some key capabilities that I hope to see in future AI are: continual learning, long-horizon memory, and goal-derived agentic behavior.

Currently, I'm interested in language models and its implications for RL (work in progress). Previously, I worked on topics in computer vision and neural fields.

Nonparametric Teaching of Implicit Neural Representations
Zhang, Chen* STS Luo*, Jason Chun Lok Li*, Yik Chung Wu, Ngai Wong
ICML, 2024
project page / code / arXiv

We showed that teaching an overparameterized MLP is consistent with teaching a nonparametric learner and proposed a sampling method that improves training time of neural fields by 30+%.

ASMR: Activation-Sharing Multi-Resolution Coordinate Networks for Efficient Inference
STS Luo*, Jason Chun Lok Li*, Le Xu, Ngai Wong
ICLR, 2024
code / arXiv

An alternative neural field architecture with near O(1) inference complexity irrespective of the number of layers. This reduces the MAC of a vanilla SIREN model by up to 500x while achieving superior reconstruction quality.

Task-Agnostic Approach to Modeling the Ventral and Dorsal Stream
STS Luo*, Tahsin Rehza*, Matthias Niemeier
MAIN, 2022
poster

Demonstrated that differences in learnt representations for classification tasks (i.e. analogous to ventral stream) and grasping tasks (i.e. analogous to dorsal strema) are driven by task requirements instead of inherited from architectural differences.

Projects

Some stuff I did that are not on the level of publications but are still quite fun.

	Something for 3D ecommerce Dorje Kongtsa, STS Luo, Marshal Guo, Jack Fan Product, In progress
	On the Effectiveness of Grid-based Neural Fields STS Luo, David Lindell Arxiv, 2024 arXiv A conjecture on the effectiveness of grid-based neural fields such as NGLOD, Instant-NGP, and DINER.
	Novel Eye-to-face Synthesis with Standard Deviation Loss Rex Tse, STS Luo, Peter Ng, Ronnie Jok Sensetime IAIF 2, 2020 paper A model that synthesizes a face from a single eye image. My first time coming up with a new loss function.
	Real-time Singing Voice Vocal Register Classification STS Luo, Justin Lam, Angel Au Sensetime IAIF 2, 2020 paper A vocal register classification model for real-time singing voice classification. My first time working with audio data.
	Novel Font Style Transfer Across Multiple Languages with Double KL-Divergence Loss Chan Lap Yan Lennon, STS Luo , Kong Chi Yui, Cheng Shing Chi Justin Sensetime IAIF 2, 2020 paper Font style transfer across multiple languages. My first time working with image style transfer.
	Cantonese Lip Reading STS Luo, Woody Lam, Julian Chu, Samuel Yau Sensetime IAIF 1, 2019 paper Created a lip reading model. My first time creating a new architecture: a pretty cool siamese model that combines CNN and LSTM.

Website design credits to Jon Barron