Title: Multi-label Learning with/without Zero-shot Classes

Abstract: Multi-label learning refers to the problem of learning to assign a subset of relevant labels to each object, drawn from a large set of candidate labels. Each object is thus associated with a binary label vector, which denotes the presence/absence of each of the candidate labels. In a real-world application, such as image/document tagging, recommender system, ad-placement, it is often the case that there is insufficient or even unavailable training data of emerging classes, which make multi-label classification even more challenging. In this talk, I cover some recent advances of multi-label learning in both Bayesian and Deep Learning, particularly those dealing with zero-short classes.

Biography: Dr. Lan Du, Data Science Lecturer in the Faculty of Information Technology and the Course Director of the Master of Data Science. He received both his B.Sc. with Honours and Ph.D. in Computer Science from the Australian National University. His research interests focus on statistical machine learning and its application in structured/unstructured data, such as text, graphs, and counts. As a Chief Investigator, he received the Google Natural Language Understanding-focused Award while he was with Macquarie University. He has published more than 40 scientific papers/articles in the top-ranked conferences and journals in machine learning, natural language processing, data mining, and general artificial intelligence. He also severed/is serving on the program committee of NeurIPS, ICML, AISTATS, AAAI, IJCAI, ACL, EMNLP, NAACL, etc.

Time: 10: 00 am, Tuesday, December 31, 2019

Venue: School of Artificial Intelligence (Room 601, Administration Building, Central Campus)

Organizer: School of Artificial Intelligence