Reporter: Zhou Tianyi
Time: November 21, 1:00 pm
Location: Management Building 304
Title: Trustworthy Multimodal Machine Learning
Abstract: Multi-modal artificial intelligence technology is being widely used in smart medical care、Unmanned systems and other important fields，Accurate design、Reliable multi-modal learning technology becomes key to supporting important applications。Multimodal data provides rich information for intelligent system decision-making，Enable multi-modal intelligent systems to “listen and understand”，Improve classification and prediction accuracy。However，In many cost-sensitive scenarios，The credibility of multi-modal fusion and decision-making is often more important。For multi-modal classification seabet sports betting tasks，Traditional methods usually assume that the quality and task correlation of each modality are stable。But actually，For different samples or in different scenarios，The quality of the modality and the relevance of the task are often dynamic。As in a multi-sensor scenario，RGB images are more effective in good light，Near-infrared images can provide more important information in difficult visual situations。This article uses uncertainty to model this dynamic、Integrating multimodal information using an improved evidence fusion strategy。The paper is titled Trusted Multi-View Classification ，Now ICLR 2021，TPAMI included。Based on this，Not only can we obtain more stable classification results when the modal quality changes dynamically，Also able to estimate decision-making confidence，And conduct modal-level traceability of classification results and classification confidence。Overall，The proposed method seabet mobile is used in multi-modal collaborative learning (listening and understanding)，Evidence estimation for different modalities (trust and evidence)，Thus supporting the reliability and stability of integration and decision-making。
Personal introduction: Dr. Zhou Tianyi,Graduated from Nanyang Technological University, Singapore, Currently working as a Principal Scientist at the Artificial Intelligence Center of the Singapore Agency for Science, Technology and Research, And serves as the leader of the artificial intelligence team (Group Manager)。 Dr. Zhou Tianyi presides over a number of key R&D projects in Singapore，And already in machine learning, Artificial Intelligence,Published more than 100 papers in core journals (Zone 1, Chinese Academy of Sciences) and international conferences (CCF Category A) in information security and other fields; Also Springer Nature Computer Science, Associate editor/invited editorial board member of important international SCI journals seabet online sports betting such as IEEE Transactions；is the co-chairman of the special report organization of several top international/important academic conferences (such as CCF Class A conference IJCAI) and the co-chairman of the technical program committee of the international flagship conference MOBIMEDIA 2020；Obtained IJCAI，ECCV，Best paper award in many top international/important academic conferences and special reports such as ACML；Served as NeurIPS, ICML, ICLR, AAAI, Area Chair (Area Chair) of top international conferences such as IJCAI。
Reporter: Chen Cen
Location: Management Building 304
Title: Co-design of high-performance algorithms and architecture for graph neural networks
Summary: Recent，Graph Neural Network (GNN) extends deep learning to the learning of graph-structured data，And demonstrated its powerful graph representation learning capabilities on many tasks。Typical graph neural network models mostly use neighborhood message propagation mechanism，Update the characteristics of the target node by seabet mobile aggregating the characteristics of neighbor nodes。Pass analysis，We found that a simple implementation of the neighborhood message propagation mechanism will lead to a large amount of redundant computation and redundant communication overhead。This report mainly introduces some of our research work on high-performance graph neural networks。We remove redundancy in the calculation process through the co-design of algorithms and architecture，Accelerate graph neural networks。Compared to current graph neural network accelerators，The graph neural network accelerator we proposed brings considerable acceleration ratio and greatly reduces energy consumption without losing network accuracy。
Personal introduction: Dr. Chen Cen，Singapore Academy of Science and Technology (A*STAR)、Information and Communications Research Institute (I2R) Research Scientist (Scientist III)，National University of Singapore、Part-time doctoral supervisor at Nanyang seabet casino review Technological University，Recipient of Singapore AI Talent Special Allowance。Main research areas: High-performance algorithms and architecture for big data and artificial intelligence。Has published more than 60 papers in international academic journals and conferences，One paper involves MICRO，HPCA, DAC, AAAI, IEEE TC,IEEE TPDS, IEEE TSMC-S, IEEE TCYB, IEEE TNNLS, ICDM, ICPP etc.。
