Research On Catastrophic Forgetting In Class-Incremental Continual Learning Model

With the prevalence of personal smart devices,an online deep learning system that can learn continually from the online data stream is widely needed.With no limit for the size of the model,regularization-based and replay-based methods show superior performance.Regularization-based methods constrain the update of some network parameters by adding extra terms in the loss function.It doesn’t need additional space to store data from previous tasks.But,when the amount of tasks is high and the similarity between tasks is low,the extra terms cannot effectively preserve learned knowledge.To overcome forgetting,Replay-based methods store a small subset of data from previous tasks and replay this stored data during the training of new tasks.Despite it needs for an additional memory buffer,it performs well with long and dissimilar tasks.However,the crucial questions for replay-based methods are what samples need to be retrieved in Memory Retrieval and what samples need to be stored in Memory Update.This paper will experiment and discuss these problems.Aiming at the problem of being unable to effectively preserve learned knowledge,when the amount of tasks is large and the similarity between tasks is low.We propose the Triple Diversity Replay Set and Gradient Regularization based Online Continual Learning method(Knowledge Regularization Replay,KRR),a regularization-based online continual learning method.It utilizes the replay set and improves the memory update method.First,it gets the classification result of training samples of each observed task.Second,the correctly and mistakenly classified training samples are input into the filter.The filter returns representative samples that will be stored in the replay set.To more effectively utilize knowledge distillation,KRR improves the distillation loss in the loss function.The model performs a regularization on distillation loss and classification loss,which lets the model learn from both new task and old tasks.Aiming at selecting and storing samples in the replay-based methods.We propose the Forgetting Constraint Memory Management and Contrastive Crossentropy Loss based Online Continual Learning method(Cross-entropy Contrastive Replay,Ce CR),a replay-based online continual learning method.It includes the Class-focused Memory Retrieval method,the Class-mean Approximation Memory Update method,and Cross-entropy Contrastive Loss.The Class-mean Approximation Memory Update method proceeds the class-level sampling without replacement to determine which classes to be retrieved.The classes that have been retrieved in the previous retrieval operations will not be retrieved repeatedly.The Class-mean Approximation Memory Update method distinguishes correctly classified training samples and mistakenly classified training samples in the replay set and selectively adds input samples to the replay set.The Cross-entropy Contrastive Loss implements the model training with a high momentum-based update to achieve effective learning.