Continual Active Learning

Active learning (AL) seeks to learn an accurate model with a minimum amount of annotation cost. It is inspired by the observation that a model can get better performance if it is allowed to choose the data points on which it is trained. For example, the learner can identify the areas of the space where it does not have enough knowledge, and query those data points which bridge its knowledge gap. Continual learning studies approaches in which the model can learn newly encountered tasks without forgetting the previously learned tasks. This is essential in real-life settings where new tasks are constantly created, e.g. sentiment analysis for new market products not observed before.

In this proposal, we investigate the learning setting where active learning meets continual learning. This is crucial as the practitioners typically do not have labeled data for a newly arrived task. The aim is to make the best use of the available annotation budget, while taking advantage of the knowledge acquired by the model when learning previously encountered tasks. We investigate two approaches to the continual active learning problem, which are related in the underlying proposed framework, and at the same time mitigate the risk by tackling the problem from two different angles. Our approaches are general, and can be applied to various modalities, including text and vision.