X-LLM: Bootstrapping Advanced Large Language Models by Treating Multi-Modalities as Foreign Languages

Title: X-LLM: Bootstrapping Advanced Large Language Models by Treating Multi-Modalities as Foreign Languages
Authors: Feilong Chen, Minglun Han, Haozhi Zhao, Qingyang Zhang, Jing Shi, Shuang Xu, Bo Xu
Published: 7th May 2023 (Sunday) @ 02:25:42
Link: http://arxiv.org/abs/2305.04160v3

Abstract

Large language models (LLMs) have demonstrated remarkable language abilities. GPT-4, based on advanced LLMs, exhibits extraordinary multimodal capabilities beyond previous visual language models. We attribute this to the use of more advanced LLMs compared with previous multimodal models. Unfortunately, the model architecture and training strategies of GPT-4 are unknown. To endow LLMs with multimodal capabilities, we propose X-LLM, which converts Multi-modalities (images, speech, videos) into foreign languages using X2L interfaces and inputs them into a large Language model (ChatGLM). Specifically, X-LLM aligns multiple frozen single-modal encoders and a frozen LLM using X2L interfaces, where X'' denotes multi-modalities such as image, speech, and videos, and L” denotes languages. X-LLM’s training consists of three stages: (1) Converting Multimodal Information: The first stage trains each X2L interface to align with its respective single-modal encoder separately to convert multimodal information into languages. (2) Aligning X2L representations with the LLM: single-modal encoders are aligned with the LLM through X2L interfaces independently. (3) Integrating multiple modalities: all single-modal encoders are aligned with the LLM through X2L interfaces to integrate multimodal capabilities into the LLM. Our experiments show that X-LLM demonstrates impressive multimodel chat abilities, sometimes exhibiting the behaviors of multimodal GPT-4 on unseen images/instructions, and yields a 84.5% relative score compared with GPT-4 on a synthetic multimodal instruction-following dataset. And we also conduct quantitative tests on using LLM for ASR and multimodal ASR, hoping to promote the era of LLM-based speech recognition.

---

Notes

• Use Q-Former for images and video
• Use C-Former - continuous integrate and fire (“CIF”) plus transformer - for speech
• For video, use pre-trained weights from image “X2L” (I2L) and continue training on video data
• Compare with LLaVa and MiniGPT-4:
  • ability to handle visual inputs from Chinese
  • quantitative tests on using X-LLM for ASR
• transferability of English image-text alignment modules - transfer from English (Indo-European) to Chinese (Sino-Tibetan)
• Open source releases:
  • instruction tuning data (Chinese)
  • codebase, model checkpoint

Related work:

• VisualBERT - image regions concatenated with text $\to$ via self-attention learn vision-text alignment
• ViLT - ViT + BERT/GPT-2/T5 - via masked language modelling (MLM), image-text matching/contrast
• CLIP - separate vision and text encoders with contrastive loss to learn alignment

X-LLM Approach

• uses ChatGLM for language decoder
  • built on top of GLM
• uses ViT-g for image and video encoder
• use Conformer for speech encoder