Future Mobile Web
Resource-Guaranteed Deep Learning
The world's demand for increasingly capable visual applications running on diverse mobile platforms such as smartphones, AR/VR headsets, and ubiquitous sensor nodes shows no sign of slowing down. We investigate visual computing algorithms as well as the underlying hardware systems that enable efficient manipulation of massive visual data such as images, videos, point clouds, and 3D graphics.
The remarkable evolution of Web technologies over the past decade means that over two-thirds of the mobile Internet traffic are contributed by Web applications. Web applications operate atop a virtual machine layer to provide portability at the expense of dramatic runtime overhead. We architect future mobile Web systems that compute faster, last longer, and consume less cellular data.
We expect deep learning (DL) to be a fundamental building block in future computing systems. One major roadblock toward that future, however, is that DL algorithms constantly operate in resource-constrained environment. We help shape the DL-enabled future by designing efficient deep learning algorithms that formally guarantee latency/energy specifications.
Robust, Explainable Deep Learning
Managed, Event-Driven Cloud Services
Just as datacenter systems suffer from the long-tail latency issue, DL systems suffer from tail accuracy, in which a few requests exhibit poor accuracy due to uncertainties and the stochastic nature of deep learning. We investigate enabling mechanisms that allow DL–driven systems to be explainable and robust against uncertainty and to guarantee that they perform as intended without causing harmful behavior.