Please note: This seminar will be given online.
Xuehai Qian
Department of Electrical and Computer Engineering, University of Southern California
Artificial intelligence (AI) and big data have greatly changed our daily life. An important type of big data is big graph data, which represents rich structures and captures the relationship between entities. Thus, graph analytics is the key component for various important real-world applications. Recently, graph learning, e.g., Graph Neural Networks (GNNs), applies machine learning on graphs and represents vertices with their embeddings, enabling the execution of graph algorithms on the embedding space. Domain-specific system and architecture provide intuitive interface for non-expert users while transparently handling system issues. However, there are often large performance gaps between graph/machine learning systems and the native algorithms.
In this talk, I will discuss several novel ideas to bridge the gap between algorithm and system architecture for graph analytics and graph learning. Our new systems and architectures provide the best of both world — providing ease of use while achieving performance close to the most efficient native algorithms. First, I explain an effective technique that enforces the precise semantics for distributed graph processing, eliminating redundant computation and communication. Second, I describe two graph mining systems that are designed for the advanced pattern decomposition algorithm and efficient distributed execution. Third, I present the instruction set and architectural extension for CPUs for sparse computation, striking a tradeoff between flexibility and specialization. Finally, I show the challenges of training deep GNNs and a new GNN training system with hybrid partition of both GNN layer and graphs.
Bio: Xuehai Qian is an assistant professor in the Department of Electrical and Computer Engineering at University of Southern California. His research group focuses on domain-specific system and architecture for graph analytics and machine learning, big data systems, and more recently, hardware security and quantum computing. He is the recipient of W. J. Poppelbaum Memorial Award at UIUC, NSF CRII and CAREER Award, and the inaugural ACSIC (American Chinese Scholar In Computing) Rising Star Award.