Using Neural Networks to Classify Galaxies Title: Rotation-invariant convolutional neural netwo

Using Neural Networks to Classify Galaxies Title: Rotation-invariant convolutional neural networks for galaxy morphology predictionAuthor: Sander Dieleman, Kyle W. Willett, Joni DambreNo two galaxies are exactly the same. They come in many different shapes, sizes, and colors. Studying these structural properties of galaxies provides a way to group similar looking galaxies together. Their morphology, the properties that describe what a galaxy looks like, is used to study their formation and evolution.Traditionally, the classification of galaxies has been performed by someone visually inspecting each image. Now, with new large sky surveys producing extremely large amounts of images in one night, using the traditional galaxy classification inspection techniques will be time consuming and not very efficient.To address this growing problem, the need for an automated classification approach that can handle large amounts of data is increasing. One solution lies in the field of machine learning. Training computers to see and analyze images would allow for more images to be classified in a timely manner. To do this, the Galaxy Zoo crowdsourcing project teamed up with Kaggle, an online predictive modeling platform, to create The Galaxy Challenge, a competition to build the best galaxy classification model.Sander Dieleman, a PhD student at Ghent University in Belgium and first author of the paper, finished in first place with his model. He presents a convolutional neural network that classifies images of galaxies based on their morphology. A convolutional neural network (CNN) is a machine learning algorithm based on biological neural networks, such as the central nervous system. It uses artificial neurons to create connections to other artificial neurons based on the information it learns from the data.The images provided by the competition hosts were separated into two groups: a training set and an evaluation set. The training set consisted of over 61,000 color images of galaxies that were annotated with morphological data. The evaluation set consisted of almost 80,000 images, but unlike the training set, had no morphological data associated with it. As their names suggest, the training set is used to train the model and the evaluation set is used to evaluate the model’s learning capabilities.The goal of the model was to determine the probability that a galaxy belongs to a particular class. There were 37 possible classes, each one representing a possible answer that corresponds to the questions asked of Galaxy Zoo participants. The list of questions presented to Galaxy Zoo participants creates a decision tree, a tree-like graph that shows all possible decisions and outcomes, and is supposed to represent all classes described in the Hubble tuning fork classification scheme.To increase the performance of the model, Dieleman used rotational symmetry, allowing the rules of the model to be applied multiple times to various versions of the images.At the end of the competition, Dieleman’s model placed first out of 326 participants with near-perfect accuracy. Analyzing the results showed that there was agreement between the Galaxy Zoo participants’ answers and the model’s output. Due to the training, the model had difficulty identifying images with morphology features that appeared infrequently. An interesting outcome pertains to images that feature black lines running across them. The black lines are caused by dead pixels in the camera, which the model doesn’t know. As a result, it tries to classify those images as disturbed galaxies. The galaxies that were easiest to classify were elliptical galaxies. Image: A schematic of the neural network architecture displaying the steps taken to reach a prediction. (Credit: Dieleman et al.) -- source link

More pics about