Accelerating Neural Architecture Search using Performance Prediction

TLDR

Neural network hyperparameter optimization and meta‑modeling are computationally expensive because they require training many model configurations. The paper aims to predict the final performance of partially trained models using frequentist regression and to employ these predictions to enable an early‑stopping method that speeds up hyperparameter optimization and meta‑modeling up to sixfold. They train standard frequentist regression models on features derived from network architectures, hyperparameters, and time‑series validation performance, and use the predictions to drive an early‑stopping scheme for hyperparameter optimization and meta‑modeling. Empirical results show the prediction models outperform Bayesian counterparts, are simpler and faster to train, accurately predict final performance across visual classification and language modeling, generalize between model classes, and the early‑stopping method achieves up to 6× speedup while maintaining optimal model selection, establishing state‑of‑the‑art accuracy and speed.

Abstract

Methods for neural network hyperparameter optimization and meta-modeling are computationally expensive due to the need to train a large number of model configurations. In this paper, we show that standard frequentist regression models can predict the final performance of partially trained model configurations using features based on network architectures, hyperparameters, and time-series validation performance data. We empirically show that our performance prediction models are much more effective than prominent Bayesian counterparts, are simpler to implement, and are faster to train. Our models can predict final performance in both visual classification and language modeling domains, are effective for predicting performance of drastically varying model architectures, and can even generalize between model classes. Using these prediction models, we also propose an early stopping method for hyperparameter optimization and meta-modeling, which obtains a speedup of a factor up to 6x in both hyperparameter optimization and meta-modeling. Finally, we empirically show that our early stopping method can be seamlessly incorporated into both reinforcement learning-based architecture selection algorithms and bandit based search methods. Through extensive experimentation, we empirically show our performance prediction models and early stopping algorithm are state-of-the-art in terms of prediction accuracy and speedup achieved while still identifying the optimal model configurations.