Alibaba’s new mPLUG-Owl upgrade has the best of both worlds, and modal collaboration enables MLLM’s new SOTA

OpenAI GPT-4V and Google Gemini have demonstrated very strong multi-modal understanding capabilities, promoting the rapid development of multi-modal large models (MLLM), and MLLM has become The hottest research direction in the industry right now.

MLLM achieves excellent instruction following ability in a variety of visual-linguistic open tasks. Although previous research on multimodal learning has shown that different modalities can collaborate and promote each other, existing MLLM research mainly focuses on improving the ability of multimodal tasks and how to balance the benefits of modal collaboration and the impact of modal interference. remains an important issue that needs to be addressed.

• Please click the following link to view the paper: https://arxiv.org/pdf/2311.04257.pdf

• Please check the following code address: https://github.com/X-PLUG/mPLUG-Owl/tree/main/mPLUG-Owl2

• ModelScope experience address: https: //modelscope.cn/studios/damo/mPLUG-Owl2/summary

• HuggingFace experience address link: https://huggingface.co/spaces/MAGAer13/mPLUG-Owl2

In response to this problem, Alibaba’s multi-modal large model mPLUG-Owl has received a major upgrade. Through modal collaboration, it simultaneously improves the performance of plain text and multi-modality, surpassing LLaVA1.5, MiniGPT4, Qwen-VL and other models, and achieves the best performance in a variety of tasks. Specifically, mPLUG-Owl2 utilizes shared functional modules to promote collaboration between different modalities and introduces a modal adaptation module to retain the characteristics of each modality. With a simple and effective design, mPLUG-Owl2 achieves the best performance in multiple fields including plain text and multi-modal tasks. The study of modal collaboration phenomena also provides inspiration for the future development of multi-modal large models

^{Figure 1 Performance comparison with existing MLLM models}

Method introduction In order to achieve the purpose of not changing the original meaning, the content needs to be rewritten into Chinese

mPLUG-Owl2 model mainly consists of three parts:

• Visual Encoder: As a visual encoder, ViT-L/14 converts the input image with a resolution of H x W into a sequence of visual tokens of H/14 x W/14 and inputs it into the Visual Abstractor.

• Visual Extractor: Extract high-level semantic features by learning a set of available queries while reducing the visual sequence length of the input language model

• Language model: LLaMA-2-7B is used as the text decoder, and the modal adaptation module shown in Figure 3 is designed.

##Figure 2 mPLUG-Owl2 model structure

In order to align the visual and For language modality, existing work usually maps visual features into the semantic space of text. However, this approach ignores the respective characteristics of visual and text information and may affect the performance of the model due to the mismatch of semantic granularity. To solve this problem, this paper proposes a modality-adaptive module (MAM) to map visual and textual features into a shared semantic space, while decoupling visual-linguistic representations to retain the unique properties of each modality. .

Figure 3 shows the schematic diagram of the modal adaptation module

shown in Figure 3 Yes, compared with the traditional Transformer, the main design of the modal adaptation module is:

• In the input and output stages of the module, LayerNorm operations are performed on the visual and language modalities respectively. to adapt to the respective feature distributions of the two modes.

• In the self-attention operation, separate key and value projection matrices are used for visual and language modalities, but a shared query projection matrix is ​​used to decouple the key and value projections. The matrix can avoid interference between the two modalities when the semantic granularity does not match.

• By sharing the same FFN, the two modalities can promote collaboration between each other

Optimize the training strategy of mPLUG-Owl2 in Figure 4

As shown in Figure 4, the training of mPLUG-Owl2 includes two stages: pre-training and instruction fine-tuning. The pre-training stage is mainly to achieve the alignment of the visual encoder and the language model. At this stage, the Visual Encoder and Visual Abstractor are trainable, and in the language model, only the visual-related model weights added by the Modality Adaptive Module are processed. renew. In the instruction fine-tuning stage, all parameters of the model are fine-tuned based on text and multi-modal instruction data (as shown in Figure 5) to improve the model's instruction following ability.

^{Figure 5 Instruction fine-tuning data used by mPLUG-Owl2}

Experiment and results

^{Figure 6 Image description and VQA task performance}

^{Figure 7 MLLM benchmark performance}

As shown in Figure 6 and Figure 7, whether it is traditional image description, VQA and other visual-language tasks, or MMBench, Q-Bench, etc. On benchmark data sets for multi-modal large models, mPLUG-Owl2 has achieved better performance than existing work.

^{Figure 8 Plain text benchmark performance}

##Figure 9 The impact of modal adaptation module on the performance of plain text tasks

In addition, in order to evaluate the impact of modal collaboration on plain text tasks, the author also tested mPLUG -Owl2 performance in natural language understanding and generation. As shown in Figure 8, mPLUG-Owl2 achieves better performance compared to other instruction-fine-tuned LLMs. Figure 9 shows the performance on the plain text task. It can be seen that since the modal adaptation module promotes modal collaboration, the model's examination and knowledge capabilities have been significantly improved. The author analyzes that this is because multi-modal collaboration enables the model to use visual information to understand concepts that are difficult to describe in language, and enhances the model's reasoning ability through the rich information in the image, and indirectly strengthens the reasoning ability of the text.

mPLUG-Owl2 demonstrates excellent multimodal understanding capabilities and successfully mitigates multimodal hallucinations. This multi-modal technology has been applied to core Tongyi products such as Tongyi Stardust and Tongyi Zhiwen, and has been verified in ModelScope and HuggingFace open demo

The above is the detailed content of Alibaba’s new mPLUG-Owl upgrade has the best of both worlds, and modal collaboration enables MLLM’s new SOTA. For more information, please follow other related articles on the PHP Chinese website!