Research on Continual Learning

Neural networks are known to suffer from the infamous issue of Catastrophic Inference when their training data shifts in distribution. Continual Learning (CL) is a branch of Machine Learning that aims at training systems that overcome this problem. As the majority of the proposed evaluation settings for CL fails at encompassing the properties of a practical scenario, we strive to address General Continual Learning (GCL), a setting in which both domain and class distributions shift either gradually or suddenly.