Kyle Hsu
徐宏愷
PhD candidate
Stanford University
kylehsu at cs dot stanford dot edu
Hi! I’m a computer science PhD candidate at Stanford University, where I work as a member of the Stanford AI Lab. I’m advised by Chelsea Finn and Jiajun Wu and affiliated with the StatsML and SVL groups. My research and education is generously supported by a Sequoia Capital Stanford Graduate Fellowship and an NSERC Postgraduate Scholarship (Doctoral). I’m a proud Taiwanese-Canadian.
My research primarily focuses on representation learning in the context of robotics, but I also nurture a broad interest in AI.
Previously, I majored in robotics as an Engineering Science undergraduate at the University of Toronto. During this time, I did research at the Vector Institute with Roger Grosse and Dan Roy. I’ve also spent time at Google Brain with Shane Gu, Berkeley AI Research with Sergey Levine, and the Max Planck Institute for Software Systems with Rupak Majumdar.
select publications
Disentanglement via Latent Quantization
Latent quantization imposes a strong inductive bias towards learning disentangled representations.
Vision-Based Manipulators Need to Also See from Their Hands
Instrumenting manipulators with hand-centric sensing facilitates out-of-distribution generalization.
Unsupervised Learning via Meta-Learning
A simple yet effective unsupervised meta-learning pipeline for image classification pre‑training.
Lazy Abstraction-Based Controller Synthesis
Efficient controller synthesis for reachability and safety specifications via on-demand abstraction construction and adaptive spatiotemporal resolution.
other publications
On the Opportunities and Risks of Foundation Models
Foundation models are pre-trained, self-supervised, large-scale, multi-modality models that exhibit emergent functionalities and homogenize deep learning models.
Differentiable Annealed Importance Sampling and the Perils of Gradient Noise
Marrying Hamiltonian Monte Carlo with annealed importance sampling results in an unbiased estimate of marginal likelihood that supports pathwise derivatives. Surprisingly, stochastic gradient annealed importance sampling turns out to be inconsistent.
On the Role of Data in PAC-Bayes Bounds
State-of-the-art nonvacuous PAC-Bayes generalization bounds for neural network image classifiers.
Unsupervised Curricula for Visual Meta-Reinforcement Learning
A meta-learning agent develops an adaptive curriculum of visuomotor tasks by deep clustering its own trajectories.
Lazy Abstraction-Based Control for Safety Specifications
Controller synthesis for safety specifications with on-demand abstraction construction.
Multi-Layered Abstraction-Based Controller Synthesis for Continuous-Time Systems
Controller synthesis made more efficient by introducing adaptive spatiotemporal granularity.
Germanium Wrap-Around Photodetectors on Silicon Photonics
Wrapping germanium photodetectors around silicon waveguides results in devices with improved optoelectronic properties.