Scholarlink

<p>In this work, we propose an algorithm-agnostic modification to existing continual learning approaches, which improves class incremental accuracy in continual image classification. First, we emphasise the distinction between the task incremental and class incremental scenarios and show that the latter poses novel challenges to the machine learning algorithms completely independent of catastrophic forgetting. Second, we note that humans learn new concepts, not in a sterile environment, but always within a wider visual context, usually unconnected to the task at hand. Consequently, we supplement the training data of an incremental learning scenario with an unlabeled stream of diverse data. This aims to improve the predictive power of each task since its elements are now seen in the context of a broad set of visual information, forcing the learner to focus on the distinctive features of target classes ignoring the specifics of neutral examples. We evaluate our solution on popular continual learning benchmarks of CIFAR10, CIFAR100, ImageNet and a&#xa0;set of fine-grained classification datasets. Our proposal improves the class-incremental performance of almost every evaluated of-the-shelf memory-less continual algorithms, and rivals the scores of several memory-based methods that incorporate small experience replay buffers.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Class Incremental Learning and Auxiliary Unlabelled Data: The Importance of Neutral Examples

  • 期刊论文

  • 发表日期：

    2026年4月2日

查看全文
  • Igor Sieradzki,
  • Łukasz Struski,
  • Igor Tomasz Podolak,
  • Romuald Aleksander Janik

摘要

In this work, we propose an algorithm-agnostic modification to existing continual learning approaches, which improves class incremental accuracy in continual image classification. First, we emphasise the distinction between the task incremental and class incremental scenarios and show that the latter poses novel challenges to the machine learning algorithms completely independent of catastrophic forgetting. Second, we note that humans learn new concepts, not in a sterile environment, but always within a wider visual context, usually unconnected to the task at hand. Consequently, we supplement the training data of an incremental learning scenario with an unlabeled stream of diverse data. This aims to improve the predictive power of each task since its elements are now seen in the context of a broad set of visual information, forcing the learner to focus on the distinctive features of target classes ignoring the specifics of neutral examples. We evaluate our solution on popular continual learning benchmarks of CIFAR10, CIFAR100, ImageNet and a set of fine-grained classification datasets. Our proposal improves the class-incremental performance of almost every evaluated of-the-shelf memory-less continual algorithms, and rivals the scores of several memory-based methods that incorporate small experience replay buffers.