The Importance of Neural Network Hyperparameters in Determining Age Inference Quality

Abstract

Abstract To estimate precise ages for large samples across the galaxy, it has become common to train machine learning models on smaller, well-characterized samples of stars and then apply them to larger samples. As part of an undergraduate course, we used this technique to train a simple neural network with varying nodes and layers using ∼11,800 ages from the upcoming APOGEE-Kepler-3 sample of stars. We find that the fraction of stars in the testing sample whose ages are recovered to better than 30% is only weakly correlated with these hyperparameters so long as the network is well fit. However, we note that it is sensitive to the chosen training sample, and that the network is susceptible to overfitting, which tends to lead to less accurate ages, particularly for the youngest and oldest stars in the sample. We provide the Jupyter notebook for this project for others wishing to do similar exercises.