MLLM Adapters: Review of VPGs and Multimodal Fusion

Table of Links

Abstract and 1 Introduction

1. Related Work

   2.1. Multimodal Learning

   2.2. Multiple Instance Learning

2. Methodology

   3.1. Preliminaries and Notations

   3.2. Relations between Attention-based VPG and MIL

   3.3. MIVPG for Multiple Visual Inputs

   3.4. Unveiling Instance Correlation in MIVPG for Enhanced Multi-instance Scenarios

3. Experiments and 4.1. General Setup

   4.2. Scenario 1: Samples with Single Image

   4.3. Scenario 2: Samples with Multiple Images, with Each Image as a General Embedding

   4.4. Scenario 3: Samples with Multiple Images, with Each Image Having Multiple Patches to be Considered and 4.5. Case Study

4. Conclusion and References

\ Supplementary Material

A. Detailed Architecture of QFormer

B. Proof of Proposition

C. More Experiments

2. Related Work

2.1. Multimodal Learning

Recently, various vision-language models (VLMs) have been proposed to enhance the fusion of text and images. For example, TCL [42] employed triplet contrastive learning to simultaneously learn from text and images. Many state-ofthe-art MLLMs have also emerged, with one major distinction lying in the design of VPGs. For instance, FROMAGe [18] and LLaVA [24] employ a straightforward linear projection as their VPGs. On the other hand, Flamingo [2] introduces the novel use of the Perceiver Resampler, incorporating cross attention and learnable query embeddings. BLIP2 [22] innovatively employs the QFormer to improve image-text alignment. Meanwhile, MiniGPT-4 [48] integrates a frozen QFormer with additional learnable layers for enhanced performance.

\ While successful in diverse tasks, current multimodal models are primarily designed under the assumption of a one-to-one relationship between texts and image inputs. In reality, the relationship between text and images can be one-to-many or many-to-many. Effectively applying multimodal models in such scenarios poses an open challenge.

\

:::info Authors:

(1) Wenliang Zhong, The University of Texas at Arlington ([email protected]);

(2) Wenyi Wu, Amazon ([email protected]);

(3) Qi Li, Amazon ([email protected]);

(4) Rob Barton, Amazon ([email protected]);

(5) Boxin Du, Amazon ([email protected]);

(6) Shioulin Sam, Amazon ([email protected]);

(7) Karim Bouyarmane, Amazon ([email protected]);

(8) Ismail Tutar, Amazon ([email protected]);

(9) Junzhou Huang, The University of Texas at Arlington ([email protected]).

:::

:::info This paper is available on arxiv under CC by 4.0 Deed (Attribution 4.0 International) license.

:::

\