時間:106年11月30日(四)14:30-15:30
地點:工程三館015室
演講者：Dr. Jilong Xue (Microsoft Research Asia)
主持人：彭文志 教授
演講題目：Optimizing Deep Learning Computation over Modern Hardware
演講大綱：
Deep learning emerges as an important new resource-intensive workload and has been successfully applied in computer vision, speech, natural language processing, and so on. Large-scale deep learning computation for model training is becoming a necessity to cope with the ever-growing data and model sizes. Deep learning computation is typically characterized by a simple tensor data abstraction to model multidimensional matrices, a data-flow graph to model computation, and iterative executions with relatively frequent synchronizations, thereby making it substantially different from Map/Reduce style distributed big data computation. This talk will first discuss the evolvement of the emerging deep learning frameworks and design choices, then introduce our recent research effort on Wolong, an optimization framework for deep learning computation. Wolong targets to build a deep learning backend that can provide automatic optimization for both scalability and efficiency of distributed and local execution. It applies RDMA-aware data-flow graph analysis to optimize the distributed execution plan to achieve the efficient communication. It also conducts automatic operator batching and kernel fusion to avoid operator scheduling overhead in local execution on GPU. Currently, Wolong can transparently improve deep learning computation performance by up to 8 times.
講者簡歷：
Jilong Xue is a researcher in System Research Group of Microsoft Research Asia (MSRA). His research focuses on building large-scale computing systems for resource-intensive workloads such as machine learning, deep learning, etc. through leveraging modern hardware including RDMA, GPU and ASIC. Recently, he is actively working in the areas of optimizing deep learning framework to bridge AI applications and diverse hardware resources, as well as building the next generation systems for future AI workloads. Jilong received his Ph.D. in Computer Science from Peking University in 2016. He had been working on large-scale graph processing system, social network security, and streaming system during his PhD study.