Preserving Generic Knowledge in Continual Learning of Vision-Language Foundation Models

A novel approach that selectively updates a sparse set of parameters in vision-language foundation models to expand their knowledge while preserving their original capabilities.