METHODS FOR CLASSIFICATION OF GALAXY TYPES USING MACHINE LEARNING
by Christopher Lee
Abstract – Lately there has been a great sharing of astronomical data, especially the images from large telescopes such as the Hubble Space Telescope. Although all galaxies were formed by gravitational pull acting on stars, each resulted in a quite different shape. Galaxy Zoo contest in Kaggle.com offered an ideal dataset to apply machine learning to the galaxy classification problem. The dataset’s labeling process was unique in that they were statistical, directly reflecting the people’s judgement. Although the labels of the galaxy images went into subcategories as well, this research focused on the first level classification between the spiral and the elliptical galaxies. The training data they offered were numerous enough that over 10,000 images of each class could be split into training and testing data, to measure the accuracy of the classifier. Variations of LeNet were chosen, to squeeze more performance from it. The resulting accuracies were within about 95-97% in agreement to the labels, which were only 80% confident of the classification themselves.
Video not available.