Online continual learning via the knowledge invariant and spread-out properties

The goal of continual learning is to provide intelligent agents that are capable continually a sequence tasks using the knowledge obtained from previous while performing well on prior tasks. However, key challenge in this paradigm catastrophic forgetting, namely adapting model new often leads severe performance degradation Current memory-based approaches show their success alleviating forgetting problem by replaying examples past when learned. these methods infeasible transfer structural i.e., similarities or dissimilarities between different instances. Furthermore, bias current and also an urgent should be solved. In work, we propose method, named Online Continual Learning via Knowledge Invariant Spread-out Properties (OCLKISP), which constrain evolution embedding features (KISP). Thus, can further inter-instance due bias. We empirically evaluate our proposed method four popular benchmarks for learning: Split CIFAR 100, SVHN, CUB200 Tiny-Image-Net. experimental results efficacy compared state-of-the-art algorithms.