Hyperparameter Optimization Is Deceiving Us, and How to Stop It
This work is significant for machine learning researchers and practitioners who rely on hyperparameter optimization to compare and evaluate algorithms, by providing a method to avoid deceptive and inconsistent conclusions.
The paper addresses the problem of inconsistent conclusions drawn from hyperparameter optimization (HPO) due to the choice of HPO configuration. It introduces epistemic hyperparameter optimization (EHPO) and a logical framework to make the process of drawing conclusions more rigorous, proving EHPO methods that are guaranteed to be defended against deception given a bounded compute time budget.
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.