RA-BLIP: Multimodal Adaptive Retrieval-Augmented Bootstrapping Language-Image Pre-training

TL;DR

This study proposes multimodal adaptive Retrieval-Augmented Bootstrapping Language-Image Pre-training (RA-BLIP), a novel retrieval-augmented framework for various MLLMs and introduces a pre-trained multimodal adaptive fusion module to achieve question text-to-multimodal retrieval and integration of multimodal knowledge by projecting visual and language modalities into a unified semantic space.

Abstract

Multimodal Large Language Models (MLLMs) have recently received substantial interest, which shows their emerging potential as general-purpose models for various vision-language tasks. MLLMs involve significant external knowledge within their parameters; however, it is challenging to continually update these models with the latest knowledge, which involves huge computational costs and poor interpretability. Retrieval augmentation techniques have proven to be effective plugins for both LLMs and MLLMs. In this study, we propose multimodal adaptive Retrieval-Augmented Bootstrapping Language-Image Pre-training (RA-BLIP), a novel retrieval-augmented framework for various MLLMs. Considering the redundant information within vision modality, we first leverage the question to instruct the extraction of visual information through interactions with one set of learnable queries, minimizing irrelevant interference during retrieval and generation. Besides, we introduce a pre-trained multimodal adaptive fusion module to achieve question text-to-multimodal retrieval and integration of multimodal knowledge by projecting visual and language modalities into a unified semantic space. Furthermore, we present an Adaptive Selection Knowledge Generation (ASKG) strategy to train the generator to autonomously discern the relevance of retrieved knowledge, which realizes excellent denoising performance. Extensive experiments on open multimodal question-answering datasets demonstrate that RA-BLIP achieves significant performance and surpasses the state-of-the-art retrieval-augmented models.

Figures (6)

• Figure 1: Illustration of different multimodal retrieval approaches. (a) Cross-modality retrieval. (b) Late-interaction retrieval. (c) RA-BLIP multimodal adaptive retrieval. For RA-BLIP, questions, documents, images, and image-text pairs are projected into a unified multimodal space.
• Figure 2: The overall workflow of RA-BLIP consists of a retrieval stage and a generation stage. We utilize multimodal encoder $f_{\theta}(\cdot)$ to project questions and multimodal knowledge into a unified semantic space to achieve question text-to-multimodal retrieval, and exclude confusing knowledge by ranking. Additionally, we employ ASKG strategy to filter out invalid knowledge, enabling precise reasoning. The parameters of LLM and image encoder are fixed, only the Q-Former and classifier are trainable.
• Figure 3: Schematic diagram of the pre-training process of multimodal adaptive fusion module.
• Figure 4: Ablation study for RA-BLIP generation and retrieval on WebQA.
• Figure 5: Influence of varying the hyperparameter $\alpha$ of generation loss.