Speaker
Description
Self-supervised learning offers a powerful way to analyse large astronomical image datasets without relying on labelled training samples. Mohale & Lochner (2024) demonstrated this by fine-tuning a ResNet-18 model with the Bootstrap Your Own Latent (BYOL) framework to produce 512-dimensional feature representations for galaxies in the Dark Energy Camera Legacy Survey (DECaLS) survey. After reducing these features via Principal Component Analysis (PCA), they used Bayesian Gaussian Mixture Models (BGMM) to test whether the representations could recover broad morphological structure without labels, demonstrating that self-supervised features contain useful information for separating galaxy types.
In this project, I reproduce their representation-learning experiment and evaluate how useful these features are for unsupervised morphology studies. I apply a more efficient clustering strategy, MiniBatchKMeans to test whether the representations can separate basic morphological classes. To measure this, I construct a high-confidence labelled sample from the Galaxy Zoo: DECaLS catalogue and compare the cluster assignments with the volunteer classifications. I further assess the structure of the feature space using Uniform Manifold Approximation and Projection (UMAP) projections and examine the stability of the cluster assignments across repeated runs.
The results show that the PCA reduced BYOL-derived features are able to distinguish between round ellipticals, spirals, and edge-on galaxies. This demonstrates that self-supervised features provide a promising and scalable basis for unsupervised galaxy morphology analysis in future large surveys.
| Apply for student award at which level: | MSc |
|---|---|
| Consent on use of personal information: Abstract Submission | Yes, I ACCEPT |