In the future, we plan to optimise these models further and apply them to new tasks, such as zero-shot learning and self-supervised learning.