|
kylehkhsu at gmail dot com I'm a visiting student researcher at Berkeley AI Research working with Sergey Levine and Chelsea Finn in the Robotic AI & Learning Lab. I'm pursuing my Bachelor's in Engineering Science (Major in Robotics Engineering) at the University of Toronto. Previously, I worked on abstraction-based controller synthesis with Rupak Majumdar at the Max Planck Institute for Software Systems. I've also spent time under Joyce Poon at the University of Toronto and Ming C. Wu at UC Berkeley. CV / Google Scholar / LinkedIn / GitHub |
|
|
I've had the fortune of participating in a diverse range of interesting research projects with great mentors and company. |
|
Kyle Hsu, Sergey Levine, Chelsea Finn NeurIPS Workshop on Meta-Learning, 2018 under review for International Conference on Learning Representations (ICLR), 2019 poster / project page / code Unsupervised learning is commonly used as a pre-training step for downstream task learning. However, the objective used during unsupervised learning is often agnostic to the downstream task type. In this work, we propose CACTUs, an unsupervised learning algorithm that leverages meta-learning techniques to learn to learn tasks constructed from unlabeled data. By incorporating knowledge of the downstream task type (image classification) into the unsupervised learning phase, CACTUs leads to significantly more effective downstream learning and enables few-shot learning without requiring labeled meta-learning datasets. This kind of unsupervised meta-learning approach is especially appealing to instantiate for domains where meta-training tasks are cumbersome to specify. |
|
*Kyle Hsu, Rupak Majumdar, Kaushik Mallik, Anne-Kathrin Schmuck under review for IEEE Transactions on Automatic Control (TAC) project page / demo / code In prior work, we introduced lazy abstraction-based controller synthesis (ABCS), but only for safety specifications. Extending lazy ABCS to handle reachability specifications requires carefully maintaining a set of frontier states from the current progress of synthesis. These frontier states are defined by their proximity to the domain of the controller. After computing the abstract transitions for frontier states, we enable synthesis to make further progress. We carefully restrict the progress synthesis can make in a single construct-then-synthesize iteration to avoid premature termination. We empirically demonstrate performance benefts of lazy multi-layered ABCS over multi-layered ABCS. |
|
*Kyle Hsu, Rupak Majumdar, Kaushik Mallik, Anne-Kathrin Schmuck IEEE Conference on Decision and Control (CDC), 2018 project page / code There is a trade-off between single- and multi-layered abstraction based controller synthesis (ABCS): the latter speeds up synthesis, but requires more computation to construct the additional coarser abstractions. This is because the abstract system computation is done in its entirety upfront, before synthesis begins. In this work, we restrict our attention to safety specifications. We extend multi-layered ABCS to be lazy: we entwine abstraction and synthesis to enable on-demand construction of the abstractions. Instead of computing all the abstractions for the entire system, the algorithm selectively chooses which portions of the system to compute abstract transitions for, avoiding doing so for portions that have been already solved by synthesis. This co-dependence of the two major components of ABCS is both conceptually appealing and results in significant performance benefits. |
|
*Kyle Hsu, Rupak Majumdar, Kaushik Mallik, Anne-Kathrin Schmuck Hybrid Systems: Computation and Control (HSCC), 2018 Abstraction-based controller synthesis (ABCS) is a class of techniques for correct-by-construction controller synthesis for nonlinear, perturbed, hybrid systems and linear temporal logic specifications. First, a finite, discrete abstraction of the original system and specification is constructed. Second, the abstract problem is solved using standard techniques from software verification. Third, the abstract solution is refined into a controller for the original system. Prior to this work, the procedure of ABCS was carried out in a single-layered setting; only one abstract system is constructed and used to find a controller. This is computationally wasteful because bottleneck regions of the system dictate the resolution (and therefore computational burden) at which ABCS operates for the entire problem. In this work, we generalize ABCS and present multi-layered ABCS: we construct multiple abstractions of the system at various spatiotemporal granularity, and do synthesis by adaptively switching between these layers, prioritizing the use of coarser, cheaper layers wherever possible, but leveraging the precision of finer, more expensive layers when necessary. We prove soundness and completeness properties of our approach, and empirically verify that it is a first step towards scalable ABCS. |
|
Ryan Going, Tae Joon Seok, Jodi Loo, Kyle Hsu, Ming C. Wu Optics Express, 2015 We present a new photodetector architecture based on coupling the silicon waveguide and germanium diode by wrapping the latter around the former instead of stacking them together. As a high school student, I contributed by testing fabricated samples. |
|
*Alphabetical authorship order. |
|
These are works that I consider particularly important to me, at least nostalgically. They are listed (roughly) in the order I read them, bottom-up. Some simply provided escapism. Many had a developmental impact on my worldview, introduced me to important concepts, or enriched my imagination. The earlier ones, in particular, are childhood favourites.
|