Welcome to Galaxy-Classification-via-Machine-Learning Discussions!

[SALMONPreet]

Discussion: Galaxy Morphology Classification with Deep Learning

Welcome to the discussion space for my project on galaxy morphology classification using Convolutional Neural Networks (CNNs).
This project builds on the Galaxy Zoo Challenge dataset, applying deep learning methods to classify galaxies into 37 morphological categories.

---

🌌 Project Overview

Galaxies come in many shapes and forms—spiral, elliptical, irregular, and more subtle categories defined by their structure.
Traditionally, classification required manual inspection by astronomers and citizen scientists. With over millions of galaxies observed, this is no longer scalable.

In this project, I:

• Implemented a ResNet-18 based CNN model to classify galaxies.
• Trained the model on ~60,000 images from the Galaxy Zoo dataset.
• Evaluated performance on a separate test set of ~70,000 images.
• Applied Grad-CAM visualizations to interpret the learned features.

---

📊 Key Features

• Custom PyTorch Dataset Loader for handling Galaxy Zoo images and labels.
• ResNet-18 architecture fine-tuned for 37 classification outputs.
• Training pipeline with optimizer, loss function, and checkpoint saving.
• Prediction script for classifying unseen galaxies.
• Grad-CAM module to visualize what parts of a galaxy the model focuses on.

---

🔬 Why This Matters

This project is part of the larger effort to bring automation and interpretability into astronomy.
Key motivations include:

• Speeding up the classification of vast sky surveys.
• Reducing dependence on manual annotation.
• Enabling astronomers to focus on rare or unusual structures.

---

💡 Discussion Points

Here are some possible directions for discussion:

1. Improving the model — would a deeper network (ResNet-50, EfficientNet) provide a significant boost?
2. Data preprocessing — what augmentations or normalization strategies are most effective for astronomical images?
3. Interpretability — how reliable is Grad-CAM in explaining CNN decisions in astronomy?
4. Applications — beyond classification, could this framework extend to anomaly detection or galaxy evolution studies?

---

📂 Next Steps

I am planning to:

• Experiment with transfer learning from other astrophysical datasets.
• Benchmark different architectures for speed and accuracy trade-offs.
• Explore semi-supervised learning to handle unlabeled galaxy data.

---

🗨️ Join the Discussion

I’d love to hear your thoughts, ideas, and critiques.

• What improvements can be made?
• How can we bridge machine learning results with real astronomical insights?
• Have you worked on a similar project? Share your experience!

Let’s collaborate to make galaxy classification more robust and insightful 🚀