Hyperparameter Optimization Is Deceiving Us, and How to Stop It

TL;DR

This paper highlights the deceptive nature of hyperparameter optimization in ML research, introduces a rigorous epistemic framework to prevent misleading conclusions, and validates a defended random search method.

Contribution

It proposes epistemic hyperparameter optimization (EHPO), a logical framework to ensure reliable conclusions, and demonstrates a defended random search method within this framework.

Findings

EHPO framework guarantees against deception within compute bounds

A defended variant of random search is proven and empirically validated

Inconsistent HPO results can be systematically addressed using the proposed framework

Abstract

Recent empirical work shows that inconsistent results based on choice of hyperparameter optimization (HPO) configuration are a widespread problem in ML research. When comparing two algorithms J and K searching one subspace can yield the conclusion that J outperforms K, whereas searching another can entail the opposite. In short, the way we choose hyperparameters can deceive us. We provide a theoretical complement to this prior work, arguing that, to avoid such deception, the process of drawing conclusions from HPO should be made more rigorous. We call this process epistemic hyperparameter optimization (EHPO), and put forth a logical framework to capture its semantics and how it can lead to inconsistent conclusions about performance. Our framework enables us to prove EHPO methods that are guaranteed to be defended against deception, given bounded compute time budget t. We demonstrate our framework's utility by proving and empirically validating a defended variant of random search.