Structural Priors in Deep Neural Networks

Yani Iannou - Microsoft Research Ph.D. Scholar, Department of Engineering, University of Cambridge

March 12, 2018, 10 a.m. - March 12, 2018, 11 a.m.




This talk will provide an overview of my PhD work, including Training CNNs with Low-rank Filters (ICLR 2016), and Deep Roots - training CNNs with sparse inter-filter connectivity (CVPR 2017). I will relate these contributions to their usage in the current state of the art networks for the Imagenet Large-Scale Visual Recognition Challenge (ILSVRC), including Inception and ResNeXt.


I'm a Microsoft Research Ph.D. Scholar in the Department of Engineering at the University of Cambridge, supervised by Professor Roberto Cipolla, head of the Computer Vision and Robotics group in the Machine Intelligence Lab, and Dr. Antonio Criminisi, a principal researcher at Microsoft Research. My PhD has focused on structural priors for deep learning, specifically the effect of encoding our prior knowledge of a problem and its representation into the structure of deep neural networks. This research has demonstrated that structural priors result in better generalization, while also improving computational efficiency.


These methods have since been embraced by state-of-the-art architectures such as Google's Inception v.3, Xception and Facebook's ResNeXt, amongst others. Our proposed structural priors are already being used in applications of deep learning for computer vision, in web-based services such as Google Photos, and on mobile phones/embedded devices, such as Apple's iPhone X.


In the past I have worked on 3D computer vision, towards methods for processing and recognizing objects in large unorganized point clouds. Outside of research, I've worked on open source projects such as the Linux kernel and the Point Cloud Library.