Hyperparameter Importance for Image Classification by Residual Neural Networks

摘要

Residual neural networks (ResNets) are among the state-of-the-art for image classification tasks. With the advent of automated machine learning (AutoML), automated hyperparameter optimization methods are by now routinely used for tuning various network types. However, in the thriving field of deep neural networks, this progress is not yet matched by equal progress on rigorous techniques that yield information beyond performance-optimizing hyperparameter settings. In this work, we aim to answer the following question: Given a residual neural network architecture, what are generally (across datasets) its most important hyperparameters? In order to answer this question, we assembled a benchmark suite containing 10 image classification datasets. For each of these datasets, we analyze which of the hyperparameters were most influential using the functional ANOVA framework. This experiment both confirmed expected patterns, and revealed new insights. With these experimental results, we aim to form a more rigorous basis for experimentation that leads to better insight towards what hyperparameters are important to make neural networks perform well.