Learning beyond Label Annotations
Date
Thursday, January 25, 2018 10:00 - 11:00
Speaker
Viktoriia Sharmanska (Universtiy of Sussex)
Location
Mondi Seminar Room 3, Central Building
Series
Seminar/Talk
Tags
Mathematics and CS Seminar
Host
Christoph Lampert
Contact
Can we incorporate annotation disagreements of the crowdsourced data collection process?Can we make use of discarded features from filter/wrapper feature selection methods? Can we build an image classifier that can incorporate knowledge from video data?Can we unify all the previous questions in a single joint learning framework?Yes, we can, the framework is called a Learning using Priviledged Information (LUPI).LUPI provides a mechanism to incorporate additional information that is only available at training time, annotation disagreements--discarded features--video dataset, into the learning process of a classifier. In this talk, I will summarize our contributions to LUPI paradigm including a Bayesian and non-Bayesian perspective of the LUPI and its variety of applications in computer vision and feature selection.