Large-Scale Adversarial Training for Vision-and-Language Representation Learning

TL;DR

VILLA introduces a large-scale adversarial training framework in the embedding space for vision-and-language models, significantly improving performance across multiple tasks by promoting invariance and robustness.

Contribution

It pioneers large-scale adversarial training in the embedding space for V+L models, combining task-agnostic pre-training and task-specific finetuning with a novel regularization approach.

Findings

Achieved new state-of-the-art results on multiple V+L benchmarks.

Demonstrated the effectiveness of embedding-space adversarial training.

Showed improved robustness and invariance in learned representations.

Abstract

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.