Supervisor

Suitable for

Abstract

Continual learning is an open problem with various real-world applications. While there is a large body of work on learning from large corpus of static well-curated data, learning from streams where the distribution of the data varies per time step is a folds more challenging problem. We seek to formulate settings where there is an adversary in the loop that can potentially alter or perturb the samples presented by the stream. The adversaries goal is break continually learning algorithms without being detected. On the other hand, efficient robust continual learners can preserve a high accuracy even under the worst case scenario where every sample presented by the stream has been adversarially manipulated. We want to formulate this new exciting problem and propose effective robust continual learning algorithms under such challenging setting.