Understanding the mechanisms of information storage and transfer in Transformer-based models is important for driving model understanding progress. Recent work has studied these mechanisms for Large Language Models (LLMs), revealing insights on how information is stored in a model's parameters and how information flows to and from these parameters in response to specific prompts. However, these studies have not yet been extended to Multi-modal Large Language Models (MLLMs). Given their expanding capabilities and real-world use, we start by studying one aspect of these models -- how MLLMs process information in a factual visual question answering task. We use a constraint-based formulation which views a visual question as having a set of visual or textual constraints that the model's generated answer must satisfy to be correct (e.g. What movie directed by the director in this photo has won a Golden Globe?). Under this setting, we contribute i) a method that extends causal information tracing from pure language to the multi-modal setting, and ii) VQA-Constraints, a test-bed of 9.7K visual questions annotated with constraints. We use these tools to study two open-source MLLMs, LLaVa and multi-modal Phi-2. Our key findings show that these MLLMs rely on MLP and self-attention blocks in much earlier layers for information storage, compared to LLMs whose mid-layer MLPs are more important. We also show that a consistent small subset of visual tokens output by the vision encoder are responsible for transferring information from the image to these causal blocks. We validate these mechanisms by introducing MultEdit, a model-editing algorithm that can correct errors and insert new long-tailed information into MLLMs by targeting these causal blocks.
AVRT: Audio-Visual Reasoning Transfer through Single-Modality Teachers
University of Tübingen, Germany(图宾根大学);MIT, Cambridge MA, USA(麻省理工学院);IBM Research, USA(IBM研究院);MIT-IBM Watson AI Lab, USA(麻省理工-IBM沃森人工智能实验室);Tuebingen AI Center, Germany(图宾根人工智能中心)
Robust Brain Tumor Segmentation with Incomplete MRI Modalities Using Hölder Divergence and Mutual Information-Enhanced Knowledge Transfer
MSU-BIT-SMBU Joint Research Center of Applied Mathematics, Shenzhen MSU-BIT University(深圳MSU-BIT大学应用数学联合研究中心);Institute of Control Theory and Control Engineering, School of Automation, Beijing Institute of Technology(北京理工大学控制理论与控制工程研究所);School of Mathematics and Statistics, Beijing Institute of Technology(北京理工大学数学与统计学学院);Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ)(广东人工智能与数字经济实验室(深圳))
CSDN: Cross-modal Shape-transfer Dual-refinement Network for Point Cloud Completion
Nanjing University of Aeronautics and Astronautics(南京航空航天大学);Delft University of Technology(代尔夫特理工大学);Lingnan University(岭南大学);The Hong Kong Polytechnic University(香港理工大学)
X-Trans2Cap: Cross-Modal Knowledge Transfer using Transformer for 3D Dense Captioning
The Chinese University of Hong Kong (Shenzhen)(香港中文大学(深圳));The Future Network of Intelligence Institute(未来网络智能研究院);Shenzhen Research Institute of Big Data(深圳大数据研究院);Shanghai Jiao Tong University(上海交通大学);Sun Yat-sen University(中山大学)
IMUGPT 2.0: Language-Based Cross Modality Transfer for Sensor-Based Human Activity Recognition
Georgia Institute of Technology(佐治亚理工学院);College of Computing, Georgia Institute of Technology(佐治亚理工学院计算学院);School of Interactive Computing, Georgia Institute of Technology(佐治亚理工学院交互计算学院);Emory University(埃默里大学)
DIGEST: Deeply supervIsed knowledGE tranSfer neTwork learning for brain tumor segmentation with incomplete multi-modal MRI scans
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences(中国科学院深圳先进技术研究院);University of Chinese Academy of Sciences(中国科学院大学);Peng Cheng Laboratory(鹏城实验室);Guangdong Provincial People’s Hospital(广东省人民医院);Guangdong Academy of Medical Sciences(广东省医学科学院)
Cross-Modal Few-Shot Learning: a Generative Transfer Learning Framework
National Engineering Research Center for Multimedia Software(国家多媒体软件工程技术研究中心);Institute of Artificial Intelligence(人工智能研究院);School of Computer Science(计算机学院);Wuhan University(武汉大学);Hubei Key Laboratory of Multimedia and Network Communication Engineering(多媒体与网络通信工程湖北省重点实验室);Centre for Frontier AI Research(前沿人工智能研究中心);Agency for Science, Technology and Research(新加坡科技研究局);University of Maryland, College Park(马里兰大学帕克分校);University of Toronto(多伦多大学);MBZUAI(穆罕默德·本·扎耶德人工智能大学)
Lightweight Facial Landmark Detection in Thermal Images via Multi-Level Cross-Modal Knowledge Transfer
Human-Robot Interfaces and Interaction Laboratory, Istituto Italiano di Tecnologia, Genoa, Italy(人机交互实验室,意大利技术研究院,热那亚,意大利);Ph.D. Program of National Interest in Robotics and Intelligent Machines (DRIM), Università di Genova(机器人与智能机器国家利益博士项目,热那亚大学)
Unsupervised Multi-modal Style Transfer for Cardiac MR Segmentation
Biomedical Image Analysis Group, Imperial College London(伦敦帝国理工学院生物医学图像分析小组);Data Science Institute, Imperial College London(伦敦帝国理工学院数据科学研究所);Department of Medicine, Imperial College London(伦敦帝国理工学院医学院)