Abstract: This talk will introduce my work on accelerating two machine learning applications on HPC systems. My parallel algorithm design focuses on the following factors: good scalability, optimized communication cost, improved or similar convergence rate, and comparable optimization cost or solution quality. I will first present how to integrate the synchronization-reducing algorithm and the sampling methods to a new parallel clustering algorithm. Later I will present a framework to solve combinatorial optimization problems over large graphs using reinforcement learning. I will also discuss different parallelism strategies and performance optimization methods over the new framework.
CS Seminar Series