Physical-layer programmability enables the network topology to be changed dynamically. In this talk, the speaker will make a case that cloud data center networks can be significantly easier to manage with physical-layer programmability. Three example network architectures will be shown as different use cases of this approach. ShareBackup enhances reliability through sharing backup switches efficiently network-wide, where a backup switch can be brought online instantaneously to recover from failures. Flat-tree solves the problem of choosing the right network topology for different cloud services by dynamically changing topological clustering characteristics of the network. OmniSwitch is a universal building block of data center networks that supports automatic device wiring and easy device maintenance. At the end of the talk, the speaker will briefly introduce an ongoing follow-up research that extends physical-layer programmability from data center networks to backbone networks.

Two important challenges in image and video understanding are designing more efficient deep Convolutional Neural Networks and learning models that are able to achieve higher-level understanding. In this talk, I will present some of my recent works towards tackling these challenges. Specifically, I will introduce the Octave Convolution [ICCV 2019], a plug-and-play replacement for the convolution operator that exploits the spatial redundancy of CNN activations and can be used without any adjustments to the network architecture. I will also present the Global Reasoning Networks [CVPR 2019], a new approach for reasoning over arbitrary sets of features of the input, by projecting them from a coordinate space into an interaction space where relational reasoning can be efficiently computed.  The two methods presented are complementary and achieve state-of-the-art performance on both image and video tasks. Aiming for higher-level understanding, I will also present our recent works on vision and language modeling, specifically our work on learning state-of-the-art image and video captioning models that are also able to better visually ground the generated sentences with [CVPR 2019] or without [arXiv 2019] explicit localization supervision. The talk will conclude with current research and a brief vision for the future.