Retrieval-based Knowledge Augmented Vision Language Pre-training

With recent progress in large-scale vision and language representation learning, Vision Language Pretraining (VLP) models have achieved promising improvements on various multi-modal downstream tasks. Albeit powerful, these pre-training models still do not take advantage of world knowledge, which is implicit in multi-modal data but comprises abundant and complementary information. In this work, we propose a REtrieval-based knowledge Augmented Vision Language Pre-training model (REAVL), which retrieves world knowledge from knowledge graphs (KGs) and incorporates them in vision-language pre-training. REAVL has two core components: a knowledge retriever that retrieves knowledge given multi-modal data, and a knowledge-augmented model that fuses multi-modal data and knowledge. By novelly unifying four knowledge-aware self-supervised tasks, REAVL promotes the mutual integration of multi-modal data and knowledge by fusing explicit knowledge with vision-language pairs for masked multi-modal data modeling and KG relational reasoning. Empirical experiments show that REAVL achieves new state-of-the-art performance uniformly on knowledge-based vision-language understanding and multimodal entity linking tasks, and competitive results on general vision-language tasks while only using 0.2% pre-training data of the best models.