表現空間における分類複雑性の評価に基づく継続学習分析手法の提案

書誌事項

タイトル別名
  • A Novel Analytical Method Based on Classification Complexity in Representation Spaces for Continual Learning

抄録

<p>A classification model must deal with incremental changes of classification tasks in practical use, which is the aim of continual learning. In particular, many existing studies focus on class incremental learning, which requires a model to learn new classes while maintaining the ability to recognize the classes the model has already learned. However, a neural network model completely forgets the ability to recognize all the learned classes when learning new classes. This phenomenon is called catastrophic forgetting and known as the main problem in continual learning. To overcome catastrophic forgetting, various continual learning methods are proposed. However, there is little understanding of their mechanism for mitigating forgetting. In this paper, we propose a novel analytical method based on classification complexity in representation spaces to reveal the properties of class incremental learning methods. To evaluate classification complexity, we design new metrics based on Local Set Cardinality average, which is the existing complexity metric. Our analytical method reveals the properties of class incremental learning methods through the evaluation of various classification complexities such as the complexity of the classification among learned classes. To verify the usefulness of our analytical method, we analyze three typical class incremental learning techniques on the two-task and five-task class incremental learning setups.</p>

収録刊行物

詳細情報 詳細情報について

問題の指摘

ページトップへ