In this talk, I discuss how we can learn deep representations that attempt to directly minimize domain shift, thereby enabling models to generalize beyond the source domain. This approach is easily extensible, can be used to align domains at both the feature and pixel level, and is widely applicable to a variety of vision tasks, including classification, detection, and segmentation.

Dissertation Talk

In this talk, I will discuss the sample complexity of reinforcement learning as gleaned from three simple settings: the linear quadratic regulator with unknown dynamics, nonlinear system identification, and multi-agent bandits in matching markets.

Machine learning systems assist in many tasks, for example, by identifying objects or people in images or by detecting spam. However, they can be vulnerable to deception. An adversary can create an image, called an adversarial example, that looks to humans like one thing but like something else to the computer. The difference from a correctly classified image must be small enough to avoid human...

Free EXPO or Conference Discount Code BSAC100

What? For three days researchers, educators, makers and enthusiasts of Snap! from all over the world will gather in the iconic Berkeley, California to share ideas, experiences, schemes and artifacts around the beauty and joy of computing. Come and join us! Engage in lively demos, discussions, talks and pitches, and prepare to mingle with actual young thinkers. Where? The Conference will take...

Code Q21

The cloud has become the most common way to deliver commercial software, but it requires building products in a very different way from traditional software, which has not been heavily studied in research. I will explain some of these challenges based on my experience at Databricks, a startup that provides a data analytics platform as a service on AWS and Azure. Databricks manages...