A Novel Class-wise Forgetting Detector in Continual Learning

Deep learning model suffers from catastrophic forgetting when learning continuously from stream data. Existing strategies for continual learning suppose the forgetting always happens when learning a new task and only deals with the previous task's global forgetting. This study introduces a novel active forgetting detector based on a windowing technique that monitors the model's forgetting rate for each encountered class label. When the model experiences the forgetting issue, we adapt the forgetting classes by using a proposed replay from experience method called online triplet rehearsal. We conduct comprehensive experiments on four vision datasets to demonstrate that the proposed approach performs significantly better than three state-of-the-art continual learning methods.