COMET: Method for Continual Vision-Language Learning
In brief
- COMET decouples distillation, routing, and capacity control for continual VLM learning.
- Tri-Affinity Distillation preserves image-image, text-text, and image-text geometry.
- A contextual bandit manages merging or retaining task-specific experts post-training.
A new approach, COMET, is introduced to enhance continual learning for vision-language models (VLMs) like CLIP. It decouples distillation, routing, and capacity control to preserve modal consistency and avoid task interference. Results on the X-TAIL benchmark show significant performance gains.
This summary gives you the essentials. The full article, its figures and quotes are on cs.AI updates on arXiv.org.
Read the original articleAI-generated summary of the source, checked by rules. Report an error