Class Incremental Learning Method Based On Feature Constraints

With the development of artificial intelligence technology,in many tasks,the current machine learning models represented by neural networks have been able to approach or even exceed human level performance.However,this model is usually a static model that cannot extend its ”function”.Whenever there is new data that needs to be trained,the neural network needs to use all the data to restart training.In the real world,this approach can become very tricky in some application scenarios.Due to storage restrictions or privacy issues,a lot of data may disappear after a given time,or even cannot be stored at all.This allows the neural network to follow adapting over time is essential.Training the original model with only new data will cause the model to largely forget old knowledge,which is called ”catastrophically forgetting.” Incremental learning technology aims to alleviate the ”catastrophic forgetting” in the continuous learning process of the model.Combining the regularization method in the existing incremental learning methods,we respectively propose a class incremental learning method suitable for neural network image classification and semantic segmentation.The specific research content is as follows:Firstly,we mathematically model the neural network model of image classification,and use the idea of recursive least squares to approximate the training process of the neural network.We propose a feature space constraint algorithm that can change the direction of parameter update in the feature space to make new data as little as possible influence on old knowledge during training.According to the training process of class incremental learning,we divide the calculations that occur during training into inter-task training and intra-task training.Inter-task training uses matrix iteration to record the changes in input data.During intra-task training,the recording matrix generated by previous inter-task training can be used to obtain the projection matrix,and the projection matrix can be used to change the gradient direction in the data learning process.In addition,in response to the imbalance in the number of new and old categories in category incremental learning,we propose adaptive loss weights to balance the ratio between the gradient generated by learning new data and the knowledge distillation gradient that protects old knowledge.Experiments have proved that this method is superior to some existing incremental learning methods without using old data.Secondly,we summarized the reasons why the category incremental learning method for image classification is not suitable for category incremental semantic segmentation tasks:one is that there are background categories in semantic segmentation,and the other is that the network architectures of two tasks are different.Therefore,we use the improved knowledge distillation loss function to calculate the loss,and use the background parameter to initialize the new category parameters,which use the background class to help learn the new class.In addition,we use the feature parameter constraint algorithm to constrain the update of some parameters of the semantic segmentation network decoder in the new data training stage,and use the feature map distillation algorithm to protect some features of the encoder.Experiments show that this method can effectively alleviate the ”catastrophic forgetting”phenomenon in class incremental semantic segmentation tasks.To sum up,the class incremental learning algorithm proposed in this article for image classification and semantic segmentation can effectively alleviate the ”catastrophic forgetting”phenomenon that appears in the continuous learning process of the model,and none of the methods in this article need to use old data in the incremental training process.