Online continual learning method strengthening decision boundaries and self-supervision
To address the issue of online continual learning in image classification,which requires adapting to new data while alleviating catastrophic forgetting,replay-based methods have demonstrated excellent performance in mitigating this problem.However,most models in these methods tend to favor learning object-agnostic solutions,which are difficult to generalize and prone to forgetting.Therefore,learning features that best represent categories is crucial for solving the catastrophic forgetting issue.This paper proposed an online continual learning approach that enhanced decision boundaries and incorporates self-supervision.Firstly,this method strengthened the decision boundaries between new classes,helping the model to classify tasks more effectively.Secondly,it employed a fused self-supervised learning approach,enabling the model to better learn repre-sentative features for each class.Through experiments comparing this method with mainstream online continual learning algo-rithms on the CIFAR-10 and CIFAR-100 public datasets,it observes that when the memory buffer M is set to 100,this method achieves an average accuracy of 60.8％and an average forgetting rate of 15.5％on CIFAR-10.When M is increased to 500,the method achieves an average accuracy of 25.9％and an average forgetting rate of 13.7％on CIFAR-100.These results vali-date that the approach of enhancing decision boundaries and incorporating self-supervision effectively alleviates catastrophic forgetting.
万方数据
维普
