<span class="var-sub_title">Morning Keynote – Azalia Mirhoseini (Google)</span> SC18 Proceedings

The International Conference for High Performance Computing, Networking, Storage, and Analysis

Machine Learning in HPC Environments

Morning Keynote – Azalia Mirhoseini (Google)

Authors: Azalia Mirhoseini (Google LLC)

Abstract: Advances in computer systems have been key to the success of Machine Learning (ML) in recent years. With the ubiquitous success of ML, it is now time for a new era where we can transform the way computer systems are built -- with learning. This talk highlights some of the challenges that modern computer systems are facing, and how we can use machine learning to address them. Specifically, it will cover various combinatorial optimization problems that appear in computational graph optimizations and then delve into some of our recent efforts at Google in addressing these problems with deep Reinforcement Learning (RL). Our results show that we can use RL-based techniques to optimize such problems without the need to characterize details of the target hardware or the computational graph. Instead, RL finds a solution by only incorporating the reward function of interest such as runtime or memory. Using the reward function, RL learns the implicit trade-offs in the underlying hardware and can achieve results that outperform traditional optimization techniques that rely on heuristics.

Back to Machine Learning in HPC Environments Archive Listing

Back to Full Workshop Archive Listing