🌌 About the Project
Galaxies come in all shapes, sizes, and colors, and they can be categorized into many different subcategories. As better and bigger telescopes continue to collect images of these galaxies, the datasets begin to explode in size.
Gaining more insight into the distribution of galaxies may lead us to better understand the evolution of the universe, therefore we come up with several questions around the topic of deep learning that may help us do so:
How does a simple Convolutional Neural Network (CNN) trained from scratch perform in comparison to more advanced models trained on the Kaggle competition “Galaxy Zoo - The Galaxy Challenge”?
How simple can the model be in order for it to have enough expressive power to achieve above 80% accuracy?
Can we achieve a better performance using transfer learning?
🎯 Methods
We hypothesize that we can construct a simple CNN architecture that outperforms a pre-trained ResNet fine-tuned on this task such that given a fixed amount of computation time, the CNN outperforms the ResNet. In a preliminary test with a simple CNN architecture, we achieved around 90% training accuracy and 84% validation accuracy on a simplified binary classification task in which the galaxies are classified into smooth galaxies and non-smooth galaxies.
🙌 Results & Final Presentation
Check out the slides to our final presentation here.
✨ Team
Sviatoslav Khukhlaev
Suraj Joshi
Alsu Surmeeva
Xueru Niu
Mathis Pink
Joseph Donovan
Nathanya Queby S.
Affiliation(s)
