Visualizing the Hidden Features of Galaxy Morphology with Machine Learning

TL;DR

This paper uses CNNs and t-SNE visualization to analyze high-dimensional features of galaxy images, revealing meaningful clustering patterns related to galaxy morphology and identifying mislabeled data.

Contribution

It introduces a method to visualize and interpret CNN-derived galaxy features, providing new insights into galaxy morphology classification.

Findings

Galaxies of the same class tend to cluster together.

Certain galaxy clusters are closer or intertwined, revealing morphological similarities.

Mislabelled galaxies can be identified through feature clustering.

Abstract

We train three convolutional neural networks (CNNs) to classify galaxies with Galaxy Zoo 2 dataset and extract the activations from the last fully connected layer or the last average pooling layer of CNNs to study the high-dimensional abstract feature representations of galaxy images. We apply t-Distributed Stochastic Neighbour Embedding (t-SNE), a popular dimensionality reduction technique, to visualize the high-dimensional galaxy feature representations in two-dimensional scatter plots. From the visualization, we try to understand the galaxy images data itself and obtain some highly valuable insights. For instance, the learned galaxy feature representations from networks indicate that the galaxies belonging to the same class tend to group together, i.e. same morphological galaxies are clustered; The cluster of completely round smooth galaxy and the cluster of in-between smooth galaxy (between completely round and cigar-shaped) are moved closer, compared to other clusters; The cluster of cigar-shaped smooth galaxy and the cluster of edge-on galaxy are intertwined surprisingly; A galaxy mislabelled as spiral galaxy in the original dataset falls in the cluster of completely round smooth galaxy, and manual inspection also identifies out the outlier as a completely round smooth galaxy. These findings will facilitate the study of galaxy morphology.