Model Agnostic Conformal Hyperparameter Optimization
share
Several novel frameworks for hyperparameter search have emerged in the last decade, but most rely on strict, often normal, distributional assumptions, limiting search model flexibility. This paper proposes a novel optimization framework based on Conformal prediction, assuming only exchangeability, and allowing for a larger choice of search model architectures and variance estimators. Several such models are explored and benchmarked against random hyperparameter search on both dense and convolutional neural networks with consistent overperformance both in final loss achieved and time to achievement.
READ FULL TEXT