Abstract

The organization of computer systems is becoming increasingly complex and heterogeneous in both software and hardware. Various hardware accelerators and device technologies breed the development of new programming models and libraries. Currently most of the programming burden is placed on programmers, who need to implement new programs or rewrite legacy code using new libraries and APIs. That often imposes a steep learning curve and non-trivial labor efforts on coding, debugging, and performance tuning. Machine learning is demonstrated to foster promising solutions to automate various aspects of coding and program optimization. This talk will discuss our efforts on making machine learning techniques feasible for addressing such real-world programming challenges with automatic program translation and optimization, including (1) a deep reinforcement learning approach to assist programming of persistent memory, which is a new memory technology recently brought to production by Intel, (2) neural-based decompilation frameworks that address several practical problems of translating binary executables to high-level code, (3) a deep symbolic superoptimization framework without requiring human knowledge, and (4) a reference identification and reasoning network for program translation.

 

Bio

Jishen Zhao is an Associate Professor in the Computer Science and Engineering Department at University of California, San Diego. Her research is at the boundary of computer architecture and systems, particularly on memory systems, machine learning for systems, and reliability. Before joining academia, she was a research scientist at HP Labs. She is a recipient of NSF CAREER award, a MICRO best paper honorable mention award, and a member of MICRO Hall of Fame.