Kyle Hsu

PhD Student

Stanford University

Hi! I’m a second-year computer science PhD student at Stanford University, where I work as a member of the Stanford Artificial Intelligence Laboratory. I’m advised by Chelsea Finn and Jiajun Wu and I’m affiliated with Stanford’s StatsML and SVL groups. My research and education is generously supported by the Stanford Graduate Fellowship.

I’m broadly interested in AI. More specifically, I care about endowing robots with key qualities of intelligent behavior. Towards this end, I enjoy thinking about a diverse array of topics, including vision, reinforcement learning, meta-learning, planning, self-supervised learning, and Bayesian inference.

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. My first research experiences were in optoelectronics and photonics under Joyce Poon at the University of Toronto and Ming C. Wu at UC Berkeley.


  • Artificial Intelligence
  • Robotics


  • BASc in Engineering Science, 2020

    University of Toronto


Differentiable Annealed Importance Sampling and the Perils of Gradient Noise

We propose an annealed importance sampling algorithm based on Hamiltonian Monte Carlo transitions that eschews the use of Metropolis-Hastings correction steps and supports the computation of pathwise derivatives. We prove both a sublinear convergence rate for Bayesian linear regression in the full-batch setting as well as an inconsistency result in the stochastic setting.

On the Role of Data in PAC-Bayes Bounds

We prove that linear PAC-Bayes bounds based on choosing the prior as the expected posterior can be improved by conditioning on a subset of the data, even with full knowledge of the underlying distribution. We apply this theoretical insight to achieve state-of-the-art, non-vacuous PAC-Bayes bounds for neural network image classifiers trained via stochastic gradient descent.

Unsupervised Curricula for Visual Meta-Reinforcement Learning

We develop an algorithm that constructs a task distribution for an unsupervised meta-learner by modeling interaction in a visual environment. The task distribution adapts as the agent explores the environment and learns to learn.

Unsupervised Learning via Meta-Learning

We propose CACTUs, a simple unsupervised learning → clustering → meta-learning pipeline for image classification pre-training. CACTUs can be thought of as a method that enables unsupervised meta-learning.

Lazy Abstraction-Based Controller Synthesis

This paper gives a self-contained presentation of lazy, multi-layered ABCS for reachability and safety specifications. It subsumes Multi-Layered Abstraction-Based Controller Synthesis for Continuous-Time Systems and Lazy Abstraction-Based Control for Safety Specifications.