ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks

TL;DR

ViLBERT introduces a multi-modal BERT-based model that learns joint visual and linguistic representations through pretraining, achieving state-of-the-art results across multiple vision-and-language tasks by treating grounding as a transferable pretraining capability.

Contribution

The paper presents ViLBERT, a novel two-stream architecture for pretraining task-agnostic visiolinguistic representations that transfer effectively to various vision-and-language tasks.

Findings

Achieves state-of-the-art results on four vision-and-language tasks.

Pretraining on Conceptual Captions improves transfer learning.

Model significantly outperforms task-specific models.

Abstract

We present ViLBERT (short for Vision-and-Language BERT), a model for learning task-agnostic joint representations of image content and natural language. We extend the popular BERT architecture to a multi-modal two-stream model, pro-cessing both visual and textual inputs in separate streams that interact through co-attentional transformer layers. We pretrain our model through two proxy tasks on the large, automatically collected Conceptual Captions dataset and then transfer it to multiple established vision-and-language tasks -- visual question answering, visual commonsense reasoning, referring expressions, and caption-based image retrieval -- by making only minor additions to the base architecture. We observe significant improvements across tasks compared to existing task-specific models -- achieving state-of-the-art on all four tasks. Our work represents a shift away from learning groundings between vision and language only as part of task training and towards treating visual grounding as a pretrainable and transferable capability.