Large-Scale Adversarial Training for Vision-and-Language Representation Learning
This work addresses the problem of improving multimodal AI systems for researchers and practitioners, representing a novel method for a known bottleneck rather than an incremental advance.
The paper tackles vision-and-language representation learning by introducing VILLA, a large-scale adversarial training method that applies perturbations in the embedding space, achieving new state-of-the-art results on multiple tasks such as Visual Question Answering and Image-Text Retrieval.
We present VILLA, the first known effort on large-scale adversarial training for vision-and-language (V+L) representation learning. VILLA consists of two training stages: (i) task-agnostic adversarial pre-training; followed by (ii) task-specific adversarial finetuning. Instead of adding adversarial perturbations on image pixels and textual tokens, we propose to perform adversarial training in the embedding space of each modality. To enable large-scale training, we adopt the "free" adversarial training strategy, and combine it with KL-divergence-based regularization to promote higher invariance in the embedding space. We apply VILLA to current best-performing V+L models, and achieve new state of the art on a wide range of tasks, including Visual Question Answering, Visual Commonsense Reasoning, Image-Text Retrieval, Referring Expression Comprehension, Visual Entailment, and NLVR2.