The paper "Semantically-Shifted Incremental Adapter-Tuning is A Continual ViTransformer" by Yuwen Tan, Qinhao Zhou, Xiang Xiang, Ke Wang, Yuchuan Wu, and Yongbin Li explores the application of pre-trained models in class-incremental learning (CIL) to enable models to continuously learn new classes while overcoming catastrophic forgetting. The authors revisit different parameter-efficient tuning (PET) methods within the context of CIL and observe that adapter tuning outperforms prompt-based methods, even without parameter expansion in each learning session. They propose incrementally tuning the shared adapter without imposing parameter update constraints, enhancing the learning capacity of the backbone. Additionally, they employ feature sampling from stored prototypes to retrain a unified classifier, further improving performance. The method estimates the semantic shift of old prototypes without access to past samples and updates stored prototypes session by session. The proposed method eliminates model expansion and avoids retaining any image samples, achieving state-of-the-art (SOTA) performance on five CIL benchmarks. The main contributions of the paper include: (1) demonstrating that incrementally tuning the adapter is a better continual learner than prompt-tuning; (2) proposing a method to retrain a unified classifier with compensated prototypes; and (3) showing superior performance on various CIL benchmarks.The paper "Semantically-Shifted Incremental Adapter-Tuning is A Continual ViTransformer" by Yuwen Tan, Qinhao Zhou, Xiang Xiang, Ke Wang, Yuchuan Wu, and Yongbin Li explores the application of pre-trained models in class-incremental learning (CIL) to enable models to continuously learn new classes while overcoming catastrophic forgetting. The authors revisit different parameter-efficient tuning (PET) methods within the context of CIL and observe that adapter tuning outperforms prompt-based methods, even without parameter expansion in each learning session. They propose incrementally tuning the shared adapter without imposing parameter update constraints, enhancing the learning capacity of the backbone. Additionally, they employ feature sampling from stored prototypes to retrain a unified classifier, further improving performance. The method estimates the semantic shift of old prototypes without access to past samples and updates stored prototypes session by session. The proposed method eliminates model expansion and avoids retaining any image samples, achieving state-of-the-art (SOTA) performance on five CIL benchmarks. The main contributions of the paper include: (1) demonstrating that incrementally tuning the adapter is a better continual learner than prompt-tuning; (2) proposing a method to retrain a unified classifier with compensated prototypes; and (3) showing superior performance on various CIL benchmarks.