A Generalized Framework for Population Based Training

TL;DR

This paper introduces a flexible, black-box Population Based Training framework that asynchronously optimizes neural network weights and hyperparameters across distributed trials, improving model performance and convergence speed.

Contribution

It presents a generalized, asynchronous PBT system that does not assume specific model architectures, supporting dynamic hyperparameter schedules and broad applicability.

Findings

Achieves better accuracy than existing methods

Demonstrates faster convergence with the same resources

Reduces sensitivity to hyperparameter choices

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. Previous PBT implementations have been synchronized glass-box systems. We propose a general, black-box PBT framework that distributes many asynchronous "trials" (a small number of training steps with warm-starting) across a cluster, coordinated by the PBT controller. The black-box design does not make assumptions on model architectures, loss functions or training procedures. Our system supports dynamic hyperparameter schedules to optimize both differentiable and non-differentiable metrics. We apply our system to train a state-of-the-art WaveNet generative model for human voice synthesis. We show that our PBT system achieves better accuracy, less sensitivity and faster convergence compared to existing methods, given the same computational resource.