Incremental learning models need to update the categories and their conceptual understanding over time. The current research has placed more emphasis on learning new categories, while another common but under-explored incremental scenario is the updating and refinement of category labels. In this paper, we present the Hierarchical Task-Incremental Learning (HTIL) problem, which emulates the human cognitive process of progressing from coarse to fine. While the model learns the fine categories, it gains a better understanding of the perception of coarse categories, thereby enhancing its ability to differentiate between previously encountered classes. Inspired by neural collapse, we propose to initialize the coarse class prototypes and update the new fine class using hierarchical relations. We conduct experiments on diverse hierarchical data benchmarks, and the experiment results show our method achieves excellent results.

增量学习模型需要随着时间的推移更新类别及其概念理解。当前的研究更加注重学习新类别，而另一个常见但尚未充分探索的增量场景是类别标签的更新和细化。在本文中，我们提出了分层任务增量学习（HTIL）问题，它模拟了人类从粗到细的认知过程。当模型学习精细类别时，它可以更好地理解粗略类别的感知，从而增强区分先前遇到的类别的能力。受神经崩溃的启发，我们建议初始化粗类原型并使用层次关系更新新的细类。