基于因果推理的长尾类增量学习  

Long-tailed Class Incremental Learning Based on Casual Inference

作　　者：刘展阳 刘进锋 LIU Zhan-yang;LIU Jin-feng(School of Information Engineering,Ningxia University,Yinchuan 750021,China)

机构地区：[1]宁夏大学信息工程学院,宁夏银川750021

出　　处：《计算机技术与发展》2025年第2期100-106,共7页Computer Technology and Development

基　　金：宁夏自然科学基金(2023AAC03126)。

摘　　要：针对传统的类增量学习方法只考虑到新任务数据的平衡分布,而忽略了现实世界中数据呈现长尾分布这一问题,提出一种因果推理方法。该方法旨在增强长尾类增量学习的性能,使模型在长尾数据分布背景下既能够逐步学习新任务的知识,又能够缓解对旧任务知识的灾难性遗忘。不同于现存的缺少基本理论的启发式方法,该方法通过追求由输入样本引起的直接因果效应,在训练中使用因果干预,在推理中使用反事实推理,以有效改善现存类增量学习模型在长尾分布数据集上的表现。与当前常用但复杂的两阶段训练方式相比,该方法更加简单有效,既不需要进行额外的微调阶段,也不需要增加额外的网络层,并且能够轻松地整合到绝大多数类增量学习方法中,从而更好地处理长尾类增量学习场景。在CIFAR-100和ImageNet-Subset数据集上进行的广泛实验验证了该方法的有效性,显示了它在各种长尾类增量学习设置中相较于之前最佳方法的优越性。Aiming at the problem that traditional class-incremental learning methods only consider the balanced distribution of data for a new task while ignoring the long-tailed distribution of data in the real world,a method based on casual inference is proposed.The method aims to enhance the performance of long-tailed class incremental learning so that the model is able to both incrementally acquire the knowledge of a new task and mitigate catastrophic forgetting of knowledge for previous tasks in the context of long-tailed data distributions.Unlike extant heuristic methods lacking fundamental theory,the proposed method uses causal intervention in training and counterfactual reasoning in inference by pursuing direct causal effects induced by input samples,in order to effectively improve the performance of existing class incremental learning models on long-tailed distributed datasets.Compared to the current commonly used but complex two-stage training approaches,the proposed method is simpler and more effective,requiring neither additional fine-tuning stages nor additional network layers.It can be easily integrated into the majority of class incremental learning methods to better handle long-tailed class-incremental learning scenarios.Extensive experiments on the CIFAR-100 and ImageNet-Subset datasets validate the effectiveness of the proposed method,showing its superiority over state-of-the-art approaches in various long-tailed class incremental learning settings.

关 键 词：因果推理 类增量学习 长尾学习 灾难性遗忘 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]