JultikaUniversity of Oulu repository

Abstract

Galaxies are systems of dark matter, stars, gas and dust orbiting around a central concentration of mass. They span a wide variety of appearances, based on which they can be classified. Several classification schemes have been developed during the last century, with an attempt to relate the different types of galaxies to each other based on their key physical features and use the received information to further the understanding of both the composition as well as the evolution of galaxies. In the past, the small amount of galaxies in survey image data allowed for the classification process to be completed visually, but with the ever-growing size and depth of survey image data, making classifications in this way nowadays is near to impossible.

Various machine learning algorithms specialized in image recognition and classification can outperform human classifiers in terms of speed and likely soon also in accuracy. The advent of the development of neural network tools for astronomy was in early 1990’s, and ever since especially convolutional neural networks have been applied to classification problems in galactic astronomy. Applications of unsupervised learning have also shown promise in being able to produce self-organizing classification results. The rapid development of machine learning algorithms and hardware suited to perform automated large-scale classification tasks with astronomical survey data holds promise for machine learning methods being able to eventually fully replace humans in most classification tasks.