TL;DR

This NeurIPS 2011 paper proposes Tree-structured Parzen Estimators (TPE) and a Gaussian-process SMBO algorithm that outperform random search and previously published human-tuned configurations on deep belief network benchmarks.

Summary

Hyperparameter tuning is a major bottleneck in training high-performing neural networks, typically left to human intuition or grid search. Two sequential model-based optimization (SMBO) algorithms are proposed — one using a Gaussian process and one using Tree-structured Parzen Estimators (TPE) — that model the conditional structure of hyperparameter spaces to identify irrelevant dimensions. Both outperform random search on deep belief network benchmarks and surpass previously published human-tuned configurations on challenging classification tasks.

Key contributions

  1. Proposes TPE, which models p(x|y) rather than p(y|x), enabling efficient conditional hyperparameter search.
  2. Introduces response surface modeling that identifies and skips irrelevant hyperparameter combinations.
  3. Demonstrates that SMBO methods outperform the best previously reported human-tuned results on DBN benchmarks.
  4. Establishes random search as a strong baseline that should be run before heavier optimization methods.

When to cite

  1. When introducing Bayesian hyperparameter optimization or citing the origin of the TPE algorithm.
  2. When arguing that conditional hyperparameter structure should be modeled rather than treating all dimensions as independent.
  3. When claiming that automated HPO can systematically outperform human-tuned configurations.
  4. When discussing the expected improvement criterion as a sample-efficient black-box optimization objective.

Machine Learning Deep Learning AI