Continual semi-supervised learning through contrastive interpolation consistency

Continual Learning (CL) investigates how to train Deep Networks on a stream of tasks without incurring forgetting. CL settings proposed in literature assume that every incoming example is paired with ground-truth annotations. However, this clashes many real-world applications: gathering labeled data, which itself tedious and expensive, becomes infeasible when data flow as stream. This work explores Semi-Supervised (CSSL): here, only small fraction input examples are shown the learner. We assess current methods (e.g.: EWC, LwF, iCaRL, ER, GDumb, DER) perform novel challenging scenario, where overfitting entangles Subsequently, we design CSSL method exploits metric learning consistency regularization leverage unlabeled while learning. show our proposal exhibits higher resilience diminishing supervision and, even more surprisingly, relying 25% suffices outperform SOTA trained under full supervision.