The article presents a novel approach in class-incremental learning named SEED. It addresses the issue of models forgetting previously acquired knowledge by selecting the most optimal expert for a task and fine-tuning it. Each expert represents each class with a Gaussian distribution, and the optimal expert is selected based on the similarity of those distributions. This approach increases the diversity and heterogeneity among the experts while maintaining the high stability of the ensemble method. The research shows that SEED achieves state-of-the-art performance in exemplar-free settings across various scenarios, demonstrating the potential of expert diversification through data in continual learning.
Publication date: 18 Jan 2024
Project Page: https://github.com/grypesc/SEED
Paper: https://arxiv.org/pdf/2401.10191
