Fei Li

DiscoVL: Unveiling Disentangled C ross-Modal Representation Learning via Orthogonal Adversarial Regularization for V ision-Language Models

Pre-trained vision-language models excel across varied perception tasks, but adapting them to novel downstream settings without sacrificing generalization remains non-trivial. Existing parameter-efficient prompt learning method often yields inconsistent representations and fails to account for semantic distribution shi...