Jump to Content

Population Based Training as a Service

Ang Li
Ola Spyra
Sagi Perel
Valentin Dalibard
Max Jaderberg
Chenjie Gu
David Budden
Tim Harley
Pramod Gupta
NIPS Systems for ML Workshop (2018) (to appear)
Google Scholar

Abstract

Population Based Training (PBT) is a recent approach that jointly optimizes neural network weights and hyperparameters which periodically copies weights of the best performers and mutates hyperparameters during training. We present in this extended abstract a generalized black-box service framework for Population Based Training. We argue that the black-box PBT service design has benefits in system scalability and required engineering effort. We perform a case study on WaveNet speech synthesis to demonstrate the effectiveness of our PBT service.