Accounting for Variance in Machine Learning Benchmarks

TL;DR

This paper models the variance in machine learning benchmarking, demonstrating how accounting for sources of variation improves comparison accuracy and proposing cost-effective methods for performance evaluation.

Contribution

It introduces a comprehensive model of benchmarking variance, analyzes comparison methods, and proposes recommendations to improve the reliability of performance assessments.

Findings

Variance from data sampling, initialization, and hyperparameters significantly affects results.

Adding more sources of variation can reduce estimation error at lower computational costs.

Proposed recommendations improve the detection of true performance improvements.

Abstract

Strong empirical evidence that one machine-learning algorithm A outperforms another one B ideally calls for multiple trials optimizing the learning pipeline over sources of variation such as data sampling, data augmentation, parameter initialization, and hyperparameters choices. This is prohibitively expensive, and corners are cut to reach conclusions. We model the whole benchmarking process, revealing that variance due to data sampling, parameter initialization and hyperparameter choice impact markedly the results. We analyze the predominant comparison methods used today in the light of this variance. We show a counter-intuitive result that adding more sources of variation to an imperfect estimator approaches better the ideal estimator at a 51 times reduction in compute cost. Building on these results, we study the error rate of detecting improvements, on five different deep-learning tasks/architectures. This study leads us to propose recommendations for performance comparisons.