1 May 2025 in IEEE Transactions on Pattern Analysis and Machine Intelligence
Recent advancements in Multimodal Large Language Models (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 scale large language or visual-language models efficiently, 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 whole 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.
15 July 2024 in IEEE Transactions on Multimedia
Current efficient approaches to building Multimodal Large Language Models (MLLMs) mainly incorporate visual information into LLMs with a simple visual mapping network such as a linear projection layer, a multilayer perceptron (MLP), or Q-former from BLIP-2. Such networks project the image feature once and do not consider the interaction between the image and the human inputs. Hence, the obtained visual information without being connected to human intention may be inadequate for LLMs to generate intention-following responses, which we refer to as static visual information. To alleviate this issue, our paper introduces LMEye, a human-like eye with a play-and plug interactive perception network, designed to enable dynamic interaction between LLMs and external visual information. It can allow the LLM to request the desired visual information aligned with various human instructions, which we term dynamic visual information acquisition. Specifically, LMEye consists of a simple visual mapping network to provide the basic perception of an image for LLMs. It also contains additional modules responsible for acquiring requests from LLMs, performing request-based visual information seeking, and transmitting the resulting interacted visual information to LLMs, respectively. In this way, LLMs act to understand the human query, deliver the corresponding request to the request-based visual information interaction module, and generate the response based on the interleaved multimodal information. We evaluate LMEye through extensive experiments on multimodal benchmarks, demonstrating that it significantly improves zero-shot performances on various multimodal tasks compared to previous methods, with fewer parameters. Moreover, we also verify its effectiveness and scalability on various language models and video understanding, respectively.
14 January 2022 in IEEE Transactions on Multimedia
Deep convolutional neural networks have achieved fairly high accuracy for single online handwritten Chinese character recognition (SOLHCCR). However, in real application scenarios, users always write multiple characters to form a complete sentence, and previous contextual information holds significant potential for improving the accuracy, robustness and efficiency of recognition. In this work, we first propose a simple and straightforward model named the vanilla compositional network (VCN) by coupling convolutional neural network with a sequence modeling architecture (i.e., a recurrent neural network or Transformer), which exploits the handwritten character's previous contextual information. Although VCN performs much better than the previous state-of-the-art SOLHCCR models, it is a two-stage architecture in nature. It suffers from high fragility when confronting with poorly written characters such as sloppy writing, and missing or broken strokes, due to relying heavily on contextual information. To improve the robustness of the OLHCCR model, we further propose a novel deep spatial & contextual information fusion network (DSCIFN). It utilizes an autoregresssive framework pre-trained on a large-scale sentence corpora as the backbone component, and highly integrates the spatial features of handwritten characters and their previous contextual information in a multi-layer fusion module. To verify the effectiveness of models, we reorganize a new form of online Chinese handwritten character with its previous context dataset, named OHCCC. Extensive experimental results demonstrate that DSCIFN achieves state-of-the-art performance and has increased strong robustness compared to VCN and previous SOLHCCR models. The in-depth empirical analysis and case study indicate that DSCIFN can significantly improve the efficiency of handwriting input because it does not need complete strokes to recognize a handwritten Chinese character precisely.