Few-shot Class-incremental Learning Based On Causality

At present,deep learning mostly uses batch learning methods based on large amounts of data.With the continuous innovation and application of artificial intelligence technology,future intelligent systems will face changing data.The purpose of few-shot class-incremental learning to enable the model to continuously iterate and learn to obtain new knowledge,thereby improving itself,improving performance and robustness to new tasks,while not forgetting old knowledge.Few-shot class-incremental learnin can not only bring machine learning closer to the level of human intelligence,but also reduce the dependence of deep learning models on a large amount of training data,making it more flexible to apply in diverse real-world environments.Therefore,few-shot class-incremental learning has gradually become the focus of researchers.In addition,most deep neural networks are generally considered "black boxes".Their parameters are obtained by automatic learning of large amounts of data,which makes it difficult to explain their internal working mechanism and judgment process.Cause-and-effect inference theory aims to improve the interpretability of deep learning models and help researchers understand the decision-making mechanisms of models.Based on this research background,this paper studies few-shot class-incremental learning based on causality from three perspectives,namely:few-shot learning,class-incremental learning,few-shot class-incremental learning,and summarizes the following contributions:1.Through the construction of a few-shot learning causal structure model,it is found that pre-training can not only enrich prior knowledge,but also confuse the final classification label.Therefore,from the perspective of causal inference,this paper uses back-door adjustment to remove the confusion of pre-training model on few-shot learning,so as to make true causal relationship between the feature and the final label.2.By observing the causal structure model in the class-incremental learning algorithm,it is found that the cause of catastrophic amnesia is that there is no link between the old data and the final classification label.Therefore,in order to alleviate catastrophic amnesia in class-incremental learning,this article adopts a class-incremental learning method based on causal effect extraction.By fixing the collision nodes in the causal structure model,this method opens up the causal path between old data and labels,achieves the causal effect of data replay,and alleviates the forgetting of old classes.3.In order to make few-shot class-incremental learning stable and interpretability,on the basis of the first two works,this article proposes a few-shot class-incremental learning for causal that combines selection strategies.First,a random scenario selection strategy is used to enhance the scalability and optimization capabilities of feature representation.Secondly,few-shot learning is applied to few-shot class-incremental learning,which makes few-shot classification robust and interpretability.Finally,from the perspective of causal inference,we use the class-incremental learning method based on causal effect in each incremental learning process to establish the relationship between old data and new labels,alleviate the catastrophic forgetting problem in few-shot class-incremental learning,and maintain the model stability.This paper presents a few-shot class-incremental learning for causal that combines selection strategies method and compares it with several mainstream few-shot class-incremental learning methods on CIFAR-100 and miniImageNet datasets.The experimental results show that the method achieves the best performance in incremental learning,which fully verifies that the algorithm proposed in this paper can effectively alleviate catastrophic forgetting and balance the stability of the model.In addition,few-shot learning based on causal intervention and causal-based incremental learning have achieved excellent performance on CUB,miniImageNet,CIFAR-100,ImageNet-Subset data,which further proves that their methods are effective and interpretability.