SONAR: Joint Architecture and System Optimization Search

TL;DR

SONAR is a joint search method that efficiently combines architecture and system optimization to find high-performance machine learning models tailored to specific hardware, significantly reducing search time.

Contribution

This paper introduces SONAR, a novel approach that interleaves architecture search and system optimization, enabling faster and more effective deployment of ML models on diverse hardware.

Findings

SONAR finds nearly optimal architectures 30 times faster than brute force.

Interleaving search processes improves efficiency and accuracy.

Early stopping enhances search speed without sacrificing quality.

Abstract

There is a growing need to deploy machine learning for different tasks on a wide array of new hardware platforms. Such deployment scenarios require tackling multiple challenges, including identifying a model architecture that can achieve a suitable predictive accuracy (architecture search), and finding an efficient implementation of the model to satisfy underlying hardware-specific systems constraints such as latency (system optimization search). Existing works treat architecture search and system optimization search as separate problems and solve them sequentially. In this paper, we instead propose to solve these problems jointly, and introduce a simple but effective baseline method called SONAR that interleaves these two search problems. SONAR aims to efficiently optimize for predictive accuracy and inference latency by applying early stopping to both search processes. Our experiments on multiple different hardware back-ends show that SONAR identifies nearly optimal architectures 30 times faster than a brute force approach.