Supervisor

Suitable for

Abstract

Data on large systems is often stream lined and multi modal, e.g., textual, images, videos, and or sound. All this data is being accumulated while jointly changing in distribution. Moreover, much of this data presented from the stream is only partially labelled. We seek to study the problem of training models on a partially labelled streams in multi-modal setting. In particular, we seek to find new effective algorithms to performing joint self-supervised continual learning on the unlabelled data while learning in supervised fashion the labelled portion of the stream.