In this talk, I will introduce the deep fusion framework for deep convolutional neural network architecture design and present two design methods. The first method aims to improve ResNets by assembling residual branches in parallel with merge and run mappings, which results in less deep but wider networks. In the second work along the path of going wider, I will introduce interleaved group convolutions, which is a drop-in replacement of regular convolutions and more efficient in using parameters and computation than regular convolutions. The proposed two networks are evaluated on image classification.
Jingdong Wang is a Senior Researcher at the Visual Computing Group, Microsoft Research Asia. His areas of interest include computer vision, multimedia, and machine learning. At present, he is mainly working on deep learning, human understanding, person re-identification, image recognition, and indexing and compact coding for large scale similarity search. He has published 100+ papers in top conferences and prestigious international journals, such as CVPR, ICCV, ACMMM, ICML, SIGIR, TPAMI, IJCV, and so on, and one book. His paper was selected into the best paper finalist at ACMMM 2015. He has shipped a dozen of technologies to Microsoft products, including Bing image search, Cognitive service, and XiaoIce Chatbot. He has served/will serve as an associate editor for TMM, an area chair in AAAI 2018, ICCV 2017, ICIP 2017, CVPR 2017, ECCV 2016, ACMMM 2015 and ICME 2015, a track chair in ICME 2012, a special session chair in ICMR 2014, and a program committee member or a reviewer in top conferences and journals, including CVPR, ICCV, ACMMM, NIPS, SIGIR, SIGGRAPH, TPAMI, IJCV.