Mining Data Hierarchies for Semi-Supervised Deep CNNs

Release Time:2019-07-15Number of visits:240

Speaker:   Prof Tao Chen

Time:       14:00-15:00, July 17

Location:  SIST 1A 200

Host:       Prof. Shenghua Gao

Abstract:

Object detection and classification has made breakthroughs along with the advance of deep CNNs in recent years. While the discriminative object features are learned via a deep CNN, there are still several issues including (1) the large intra-class variation and deformation, (2) the availability of large-scale annotated data and (3) the uneven separability of different data categories which have become three major impediments of deep learning for image classification. We thus propose to mine the data hierarchy residing in the large-scale data itself for semi-supervised deep CNN learning. In particular, we propose a subcategory-aware CNN (S-CNN) to solve the object intra-class variation problem. and a semi-supervised hierarchical convolutional neural network (SS-HCNN) to address the large-scale data annotation and uneven data separability problem. The proposed methods have been evaluated on several state-of-the-art datasets and the results well validate the effectiveness of the proposed methods.

Bio:

陈涛博士于2008年毕业于浙江大学信电系获得硕士学位,2012年毕业于新加坡南洋理工大学获得博士学位,2018以海外高层次人才身份引进到复旦大学担任正高级研究员,博士生导师。回国前曾经先后在新加坡科技局、新加坡智能机器人实验室、新加坡资讯通信研究院等顶级研究机构从事视觉计算与机器学习的研究开发工作,承接并参与了一系列新加坡政府和企业的重大课题。此外,陈涛博士与2017年至2019年还曾在华为亚太研究院新加坡研究所从事过AI芯片的研发工作,并有多款实际芯片产品落地。迄今为止,陈博士已经在CCF A类或者JCR 一区刊物如IEEE T-PAMI/T-IP/T-CYB等期刊上发表学术论文40余篇,授权专利一项。

SIST-Seminar 18189