摘要 Abstract 

深度學習模型通常使用固定資料集進行訓練，因而無法因應資料隨著時間增加的情境。持續學習旨在模擬人類學習，隨著時間的推移不斷增加知識。此演講將概述持續學習，介紹持續學習實例分割，並探討未來研究方向。

Recent deep learning models excel in computer vision tasks, including image segmentation. However, these models are typically trained offline with fixed datasets, limiting adaptability to new data in real-world scenarios. Continual learning (CL) in deep models mimics human learning, enhancing skills and knowledge over time. This talk provides a CL overview and presents our two recent studies. The first introduces a segmentation model that continually learns new cell types while retaining knowledge of old ones in microscopy images. The second presents a CL model for instance segmentation with image-level weak supervision. The conclusion explores potential research directions in continual learning.