Continual semi-supervised learning through contrastive interpolation consistency#

This work is the result of my MSc thesis.

Existing Continual Learning methods often require labeled data for each task, which is impractical in real-world scenarios. This study explores Continual Semi-Supervised Learning (CSSL), where only a small portion of labeled data is available. Evaluating current CL methods in this setting reveals challenges of overfitting. We propose a novel CSSL method: Contrastive Continual Interpolation Consistency (CCIC), which imposes consistency among augmented and interpolated examples while exploiting secondhand information peculiar to the Class-Incremental setting. Results demonstrate its robustness to limited supervision, outperforming state-of-the-art methods when these are trained with full supervision.