Modelling Delayed Labels in Online Continual Learning

Supervisor

Suitable for

Abstract

Co-supervised by Phillip Torr and Adel Bibi from the Department of Engineering Science.  Online continual learning is the problem of predicting every sample in the stream while simultaneously learning from it. That is to say, the stream first presents data to be predicted and then the stream reveals the labels for the model to train on. However, in most real-case scenarios, labelling is an expensive laborious and time-consuming procedure. Thereof, we seek to study the sensitivity of existing continual learning algorithms when labels of images at step t are only provided at step t + k. This setting poses two challenges. (1) Learning from unlabelled data. (2) Modelling the delayed labels. To that end, we are interested in proposing new algorithms that per time step t can perform self-supervision continually while jointly training on the labelled data revealed from step t-k.