Assessor-guided learning for continual environments

This paper proposes an assessor-guided learning strategy for continual where assessor guides the process of a base learner by controlling direction and pace thus allowing efficient new environments while protecting against catastrophic interference problem. The is trained in meta-learning manner with meta-objective to boost learner. It performs soft-weighting mechanism every sample accepting positive samples rejecting negative samples. training objective minimize meta-weighted combination cross entropy loss function, dark experience replay (DER) function knowledge distillation whose interactions are controlled such way attain improved performance. A compensated over-sampling (COS) developed overcome class imbalanced problem episodic memory due limited budgets. Our approach, Assessor-Guided Learning Approach (AGLA), has been evaluated class-incremental task-incremental problems. AGLA achieves performances compared its competitors theoretical analysis COS offered. Source codes AGLA, baseline algorithms experimental logs shared publicly https://github.com/anwarmaxsum/AGLA further study.