Enhancing Task Identification Through Pseudo-OOD Features for Class-Incremental Learning

Weizhuo Zhang, Jiankang Chen, Wentao Zhang, Zhijun Tan, Ruixuan Wang

Abstract: The objective of exemplar-free class-incremental learning is to enable the model to continuously learn new classes without accessing samples from previously learned classes, which is a challenging scenario as it necessitates preventing the forgetting of knowledge related to the old classes while ensuring the ability to learn new classes. Existing approaches either advocate for the utilization of unique models for each task or heavily rely on stored relevant information from previous categories. Our method employs a shared feature extractor while training a unique classifier head for each task. In particular, we introduce task-specific adapters into the pre-trained model to facilitate the acquisition of task-specific representations. And just based on the in-distribution (ID) training data of the current task, a pseudo-OOD feature generator is proposed to assist each classifier head in developing out-of-distribution (OOD) detection ability for distinguishing samples from other tasks. This helps the model more accurately select the appropriate task-specific classifier head during inference, consequently enhancing its effectiveness in performing the class-incremental learning task. Experimental evaluations on two challenging benchmarks with various incremental settings confirm the superior performance of our method. Source code is available at https://openi.pcl.ac.cn/OpenMedIA/PseudoOODCIL.