Uni-MoE: Scaling Unified Multimodal LLMs with Mixture of Experts
(2405.11273)Abstract
Recent advancements in Multimodal LLMs (MLLMs) underscore the significance of scalable models and data to boost performance, yet this often incurs substantial computational costs. Although the Mixture of Experts (MoE) architecture has been employed to efficiently scale large language and image-text models, these efforts typically involve fewer experts and limited modalities. To address this, our work presents the pioneering attempt to develop a unified MLLM with the MoE architecture, named Uni-MoE that can handle a wide array of modalities. Specifically, it features modality-specific encoders with connectors for a unified multimodal representation. We also implement a sparse MoE architecture within the LLMs to enable efficient training and inference through modality-level data parallelism and expert-level model parallelism. To enhance the multi-expert collaboration and generalization, we present a progressive training strategy: 1) Cross-modality alignment using various connectors with different cross-modality data, 2) Training modality-specific experts with cross-modality instruction data to activate experts' preferences, and 3) Tuning the Uni-MoE framework utilizing Low-Rank Adaptation (LoRA) on mixed multimodal instruction data. We evaluate the instruction-tuned Uni-MoE on a comprehensive set of multimodal datasets. The extensive experimental results demonstrate Uni-MoE's principal advantage of significantly reducing performance bias in handling mixed multimodal datasets, alongside improved multi-expert collaboration and generalization. Our findings highlight the substantial potential of MoE frameworks in advancing MLLMs and the code is available at https://github.com/HITsz-TMG/UMOE-Scaling-Unified-Multimodal-LLMs.
Overview
-
Uni-MoE is a Multimodal Large Language Model that uses Mixture of Experts (MoE) to scale efficiently across various data modalities including images, audio, and video without significant computational costs.
-
The model architecture involves modality-specific encoders, connectors for a unified language space, and sparse MoE layers that selectively activate a few experts per input to enhance efficiency.
-
Experimental results demonstrate Uni-MoE's superior performance across tasks involving speech, audio, and image-text understanding, showcasing its versatility and reduced computational overhead.
Scaling Multimodal Models with Mixture of Experts: An Exploration of Uni-MoE
Background and Motivation
When it comes to Multimodal LLMs (MLLMs), one key question has always loomed large: how do we scale these models efficiently without ramping up computational costs? Recent advancements in this space have shown that larger models and more data significantly boost performance, but they come with a challenge—huge computational overhead during training and inference.
Traditionally, LLMs and image-text models have been scaled using the Mixture of Experts (MoE) architecture. However, the application of MoE to a unified multimodal setting—one that goes beyond just the image and text modalities to include audio, speech, video, and more—had remained relatively unexplored. That's where Uni-MoE steps in as a pioneering attempt to address this gap.
The Architecture: How Uni-MoE Works
Uni-MoE leverages modality-specific encoders combined with connectors that map these inputs into a unified language representation space. The MoE architecture is incorporated into the foundation of this model, enabling efficient training and inference. Here's a simplified breakdown of its components:
- Modality-Specific Encoders: These specialized modules process data inputs (like images, audio clips, videos, etc.) and encode them.
- Connectors: These transformers help map different modality encodings into a common representation space suitable for a language model.
- Sparse MoE Layers: Within the core language model, sparse routing mechanisms ensure that only a handful of experts are activated per input, making the process more efficient.
Training Strategy
The training of Uni-MoE is pretty novel. It follows a three-stage process:
- Cross-Modality Alignment: This stage focuses on training the connectors to map different modalities into a unified language space.
- Specialized Expert Training: Each modality-specific expert is trained on relevant cross-modal datasets to refine its capabilities.
- Unified MoE Training: Finally, the MoE and sparse routing mechanisms are fine-tuned using mixed multimodal instructional datasets with Low-Rank Adaptation (LoRA) techniques.
Experimental Results and Key Insights
Uni-MoE's effectiveness was put to the test across various multi-modal datasets, encompassing tasks related to images, videos, audio, and long-form speech understanding. Here are some headline results:
- Speech-Image Understanding: Uni-MoE achieved notable improvements over existing models like Macaw-LLM and X-InstructBLIP, especially in long speech understanding tasks, highlighting its robustness.
- Audio-Text Tasks: It surpassed state-of-the-art models by a substantial margin in tasks like ClothoAQA, demonstrating its capability to handle audio data.
- Image-Text Understanding: While dense models often perform well on image-text tasks, Uni-MoE still holds its own, delivering competitive performance.
Implications
The potential implications of Uni-MoE are noteworthy:
- Efficiency: By incorporating sparse routing, Uni-MoE significantly reduces the computational overhead, making it feasible to scale up multimodal models without hitting performance bottlenecks.
- Versatility: The ability to handle a range of modalities efficiently positions Uni-MoE as a flexible foundation for future multimodal models.
- Stability and Generalization: The novel training strategy ensures that Uni-MoE exhibits reduced performance biases across different tasks, improving the overall stability and robustness.
Future Directions
The introduction of Uni-MoE opens up several avenues for future research:
- Scaling Up: Exploring the addition of more experts and further optimizing the routing mechanisms could push the performance boundaries even higher.
- Extended Modalities: Incorporating more varied data types, including potentially more complex auditory and visual data, could further broaden the scope of Uni-MoE.
- Real-world Applications: With its efficient training and inference capabilities, deploying Uni-MoE in real-world applications such as voice-activated assistants, automated video analysis, and more could be a game-changer.
In essence, Uni-MoE represents a commendable step forward in the realm of multimodal LLMs. By addressing the intrinsic challenges of scaling and efficiency, it paves the way for future innovations in combining diverse data into cohesive and powerful AI models. Check out the code and delve deeper into Uni-MoE's unique architecture here.