Deep Learning Methodologies and Tools for Scientific Problems: Berkeley Fluids Seminar

Seminar | September 16 | 12-1 p.m. | 3110 Etcheverry Hall

 Chiyu “Max” Jiang, Ph.D. Student, Department of Mechanical Engineering, University of California, Berkeley

 Department of Mechanical Engineering (ME)

Abstract: Modern tools in machine learning and deep learning can offer new insights to old problems, both on the methodological side as well as on the technical side. The first half of this talk will present an overview of tools and subfields in geo¬metric deep learning that are particularly relevant and applicable to physical problems, (e.g., learning on regular grids, meshes/graphs and point clouds, equivalent representations using spherical harmonics etc.), including some of Jiang’s recent work in this area. Specifically for a fluids audience, Jiang will also discuss work that uses deep learning methodologies towards fluids applications. The second half of the talk will discuss various computational tools that have been developed in the deep learning realm including GPU-based frameworks (e.g., PyTorch, Tensorflow) that have various unique properties (python friendly, GPU ready, high performance, built-in differentiability). Such frameworks offer the possibility of writing high-performance linear algebra code with a high level user friendly numpy-like programming interface, that can execute computation agnostic to the specific hardware (CPU, GPU, parallel computation on a cluster). Though designed with deep learning applications in mind, these computational tools might appeal to a wider audience in other computational fields for its flexibility and high performance.

Biography: Chiyu “Max” Jiang is a 5th-Year PhD student in Mechanical Engineering advised by Dr. Philip Marcus. He is currently interning in the Perception Team at Google AI, and is affiliated with the data analytics group at Lawrence Berkeley National Lab. His research interest is in 3-dimensional machine learning / deep learning, and its applications ranging from computer vision to physics and climate science., 510-642-5942