Compensating Distribution Drifts in Class-incremental Learning of Pre-trained Vision Transformers

The exploration of class-incremental learning (CIL) in pre-trained vision transformers (ViTs) reveals critical challenges, particularly concerning distribution drift during sequential fine-tuning. The introduction of Sequential Learning with Drift Compensation (SLDC) addresses these issues by aligning feature distributions, which is crucial for maintaining classifier performance. This aligns with findings in related works, such as 'FedeCouple,' which emphasizes balancing global generalization and local adaptability in federated learning, and 'Difference Vector Equalization,' which focuses on robust fine-tuning of vision-language models. Both underscore the importance of effective distribution management in machine learning, highlighting a broader trend towards enhancing model adaptability and performance across diverse tasks.