Efficient hyperparameter optimization in convolutional neural networks by learning curves prediction

Abstract

In this work, we present an automatic framework for hyperparameter selection in Convolutional Neural Networks. In order to achieve fast evaluation of several hyperparameter combinations, prediction of learning curves using non-parametric regression models is applied. Considering that “trend” is the most important feature in any learning curve, our prediction method is focused on trend detection. Results show that our forecasting method is able to catch a complete behavior of future iterations in the learning process.