Dr.Meng Lei
NUS-Tsinghua Extreme Search Center
时间:2019年9月28日10:00-11:00
地点:浙江大学玉泉校区曹光彪主楼218室
Abstract:
Visual food logging, usually embedded in mobile phone apps, is an emerging tool for diet management. It allows users to upload food photos of their daily intake and provides personalized services to encourage users to retain a healthy eating style. To achieve these, food recognition and recommendation are two key functionalities. However, the performance of learning to recognize food content, such as food name and its ingredients, from images is usually limited by the diverse appearances of images. This also makes modeling users’ eating preferences based on these images more difficult. This talk presents our recent achievements on learning representations of food images for improved recognition and recommendation. Both are achieved by leveraging another view of food, i.e. the tagged ingredients, to regularize the encoding of image features. In food recognition, the multimodal assumption allows the use of transfer learning to map the representations of images to those of ingredients, thus taking advantage of their stronger discriminative power. Food recommendation is more challenging since users typically eat food in different categories, requiring the image features to go beyond semantics, referred to as collaborative similarity. We will show how to encode both the semantic and collaborative similarities in image representation via a continual multitask learning approach. Besides the technical details, backgrounds, key challenges, and the experimental findings will be discussed.
BIOGRAPHY:
Dr. Meng is Senior Research Fellow with NUS-Tsinghua Extreme Search Center (NExT++), working with Prof. Tat-Seng Chua. Prior that, he was with Joint NTU-UBC Research Center of Excellence in Active Living for the Elderly (LILY), Nanyang Technological University, Singapore, working with Prof. Chunyan Miao with Nanyang Technological University and Prof. Cyril Leung with the University of British Columbia.
His current research focus is on unsupervised learning, deep learning, and multimodal data fusion for applications in data mining, multimedia, and computational healthcare. He has published a book with Springer and 16 conference and journal papers at well-known and top venues, such as TKDE, IEEE Transactions on Cybernetics, TNNLS, Neural Networks, MM, SDM, ICMR, and IJCNN. He is the editorial board member of Applied Soft Computing, and has served as Program/Technical Committee member for a number of international conferences. He also frequently serves as Reviewer for a number of high-quality conferences and journals, such as MM, SIGIR, KDD, ICDM, SDM, and TNNLS. He is a member of the INNS and the IEEE.