计算机科学与探索 ›› 2024, Vol. 18 ›› Issue (2): 477-485.DOI: 10.3778/j.issn.1673-9418.2212055

• 人工智能·模式识别 • 上一篇    下一篇

融合平衡权重和自监督的类增量学习方法

巩佳义,许鑫磊,肖婷,王喆   

  1. 华东理工大学 信息科学与工程学院,上海 200237
  • 出版日期:2024-02-01 发布日期:2024-02-01

Class Incremental Learning Method Integrating Balance Weight and Self-supervision

GONG Jiayi, XU Xinlei, XIAO Ting, WANG Zhe   

  1. School of Computer Science and Engineering, East China University of Science and Technology, Shanghai 200237, China
  • Online:2024-02-01 Published:2024-02-01

摘要: 针对图像分类中类增量学习的知识灾难性遗忘现象,现有类增量学习方法着重于模型分类层的不平衡偏移修正,忽视了模型特征层的偏移,未能良好解决类增量学习所面临的新旧类别不平衡问题。为此,提出一种新的类增量学习方法,称为融合平衡权重和自监督的类增量学习方法(BWSS)。首先,BWSS利用旧类在训练中预测期望低的特点设计了自适应的平衡权重,扩大旧类的损失回传占比以修正整个模型的偏移。其次,BWSS引入自监督模块,将样例的旋转角度预测作为辅助任务,强化模型对冗余特征和共性特征的表达能力,以更好地支撑增量任务。通过与主流类增量学习算法在公开数据集CIFAR-10和CIFAR-100上的实验对比,证明BWSS不仅在类别少样例多的CIFAR-10上增量性能更优,在类别多样本少的CIFAR-100上同样具有优势。通过消融实验和特征可视化,验证了所提方法对模型的特征表示能力和增量性能是有效的。BWSS在CIFAR-10上的5阶段增量任务最终准确率达到了76.9%,相比基线方法提高了5个百分点。

关键词: 增量学习, 自监督, 不平衡问题, 知识蒸馏

Abstract: In view of the catastrophic forgetting phenomenon of knowledge in class incremental learning in image classification, the existing class incremental learning methods focus on the correction of the unbalanced offset of the model classification layer, ignoring the offset of the model feature layer, and fail to solve the problem of the imbalance between the new and old samples faced by class incremental learning. Therefore, a new class incremental learning method is proposed, which is called balance weight and self-supervision (BWSS). BWSS designs an adaptive balance weight based on the low expectation of the old class in training, so as to expand the loss return proportion of the old class in the same data batch to correct the overall model offset. Then, BWSS introduces self-supervised learning to predict the rotation angle of the sample as an auxiliary task, so as to make the model have the expression ability of redundant features and common features to better support incremental tasks. Through the experimental comparison with the mainstream incremental class learning algorithms on the open datasets CIFAR-10 and CIFAR-100, it is proven that BWSS not only has better incremental performance on CIFAR-10 with fewer categories and more samples, but also has advantages on CIFAR-100 with more categories and fewer samples. Ablation experiments and feature visualization demonstrate that the proposed method is effective for the feature representation and incremental performance of the model. The final accuracy of BWSS’s 5-stage incremental task on CIFAR-10 reaches 76.9%, which is 5 percentage points higher than the baseline method.

Key words: incremental learning, self-supervision, problem of imbalance, distillation of knowledge