Kyle Hsu

aka 徐銘謙, 徐宏愷

kylehkhsu at gmail dot com

I'm currently an undergraduate researcher at the Vector Institute working with Roger Grosse. I'll graduate with a bachelor's in Engineering Science as a robotics engineering major from the University of Toronto in June 2020.

Previously, I was a visiting student researcher at Berkeley AI Research, where I worked with Sergey Levine and Chelsea Finn on unsupervised meta-learning as a member of the Robotic AI & Learning Lab. Prior to that, I spent a summer in Germany as a DAAD RISE Scholar working on scalable abstraction-based controller synthesis with Rupak Majumdar at the Max Planck Institute for Software Systems. My first research experiences were in optoelectronics and photonics under Joyce Poon at the University of Toronto and Ming C. Wu at UC Berkeley.

I'm broadly interested in the problem of creating machines that exhibit intelligence, the hallmarks of which I consider to be adaptability, flexibility, and generality. In my exploration of this interest, I have studied and done research in program synthesis, automatic control, meta-learning, unsupervised learning, reinforcement learning, and neural network optimization.

Curriculum Vitae  /  Google Scholar  /  LinkedIn  /  GitHub

I've had the fortune of participating in a diverse range of interesting research projects with talented and patient collaborators.

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 may be especially appealing in domains where meta-training tasks are cumbersome to specify.

Lazy abstraction-based controller synthesis (ABCS) entwines abstraction (computing a finite-state, sampled overapproximation of a hybrid system) and synthesis (solving the game, defined by a control specification, in the abstraction) to increase the scalability of ABCS while preserving soundness and relative completeness. This co-dependence of the two major components of ABCS is both conceptually appealing and results in significant performance benefits over previous methods, which require upfront computation of all abstractions for the entire system. We achieve lazy abstraction by i) proving that, to make further progress at any stage of synthesis, only a frontier of states needs to be investigated at higher spatiotemporal precision; and ii) proposing efficient algorithms to compute the frontier. This paper gives a self-contained presentation of lazy, multi-layered ABCS for reachability and safety specifications.

In prior work, we introduced multi-layered abstraction based controller synthesis (ABCS). There is a trade-off between single- and multi-layered ABCS: the latter speeds up synthesis, but requires more computation to construct the additional coarser abstractions. This is because algorithms in prior work require all layers of abstraction to be computed upfront and in their entirety before the start of synthesis. In this work, we restrict our attention to safety specifications, and extend multi-layered ABCS to be lazy: we entwine abstraction and synthesis to enable on-demand construction of the abstractions. We prove soundness and relative completeness, and empirically find that lazy ABCS can result in significant computational savings over non-lazy ABCS by avoiding computing transitions for significant portions of finer, more expensive abstractions.

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. It works as follows: first, a finite, discrete abstraction of the original system and specification is constructed; second, the abstract problem is solved using standard fixed-point iteration techniques from software verification; third, the abstract solution is refined into a controller for the original system. In prior work, ABCS is usually 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 relative completeness properties of our approach, and empirically verify that it is a first step towards scalable ABCS.

We present a new photodetector architecture based on coupling the silicon waveguide and germanium diode by wrapping the latter around the former on four faces instead of one. As a high school student, I collected data by testing fabricated samples.

^†alphabetical authorship order

This guy makes a nice website.