Galaxy Morphology Classification & Representation Learning

This repository contains experiments using deep learning for galaxy image analysis.

---

Dataset

• ~8,000 RGB galaxy images
• Resolution 64×64×3
• Classes include elliptical, barred/unbarred spiral, tight/loose spirals, etc.
• High visual similarity across subclasses makes classification challenging

---

1) CNN Classifier

Architecture Overview

• Kernel size kept small (3×3) to avoid oversmoothing on small 64×64 inputs

• Progressive filter schedule: 32 → 64 → 128

• Block structure used: Conv2D → BatchNorm → ReLU → MaxPool(2×2)

• Avoided Flatten() because dense head becomes too large
  (Flattening → ~1M parameters → heavy overfitting risk)

Regularization Techniques Applied

Method	Effect
Input noise	Reduced validation loss significantly
Dropout in dense layers	Prevented memorization
L2 regularization	Helped generalize better
Label smoothing	Improved soft decision boundary
Batch size increase (32→64)	Higher validation accuracy

Performance

• Test Accuracy: ~0.82
• Frequent confusion between class 5 (Barred Spiral) and 6 (Unbarred Tight Spiral)

---

2) Variational Autoencoder + Normalizing Flows

Goal: Improve latent representation quality for realistic galaxy generation.
Normalizing flows were added to map a simple Gaussian prior into a more flexible latent distribution.

Important Training Notes

• Added β-VAE KL weighting:
  beta * KL_loss helps avoid the latent collapse → model captures colour gradients better
  This reduces the tendency for decoded outputs to drift toward grey/0.5 uniform values.
  The encoder no longer falls into the "neutral grey" minimum → noticeably richer colour reconstruction.

Sampling Code

z0 = tf.random.normal((num_samples, LATENT_DIM))  # Base Gaussian
zK, _ = flow(z0, training=False)                  # Flow transforms latent
images = vae.decoder(zK)                          # Decode to galaxy images

Step 1: Sample z0 from a standard Gaussian

The VAE assumes the latent prior is N(0, I). So generation starts by drawing noise from this base distribution:

z0 = tf.random.normal(...)

This is NOT yet the latent the model expects, it's just the unwarped space.

Step 2: Push that Gaussian sample through the Normalizing Flow:

A normalizing flow learns an invertible transformation that reshapes simple Gaussian noise into the true data latent distribution.

zK, logdet = flow(z0, training=False)

Meaning:

z0  →  z1  →  z2  →  ...  →  zK  = learned latent space

• The VAE encoder learns one approximation of the latent space,
• The flow further warps and densifies it into a more realistic manifold.

Results: Sharper details, improved colour representation, and more diverse morphologies.

Why training=False is important

Because at inference we want a deterministic mapping.

If training=True	If training=False
Flow still tries to learn	Flow uses what it learned
Latent distribution shifts	Latent stays correct
Images look worse / unstable	Images are clean + realistic

So for sampling we freeze everything → only use what was learned.

---

3) Convolutional Autoencoder (Anomaly Detection)

Used to reconstruct galaxies and detect anomalies via reconstruction error.

Preprocessing

images = np.arcsinh(images / 20.0)
channel_max = images.reshape(-1, 3).max(axis=0)
images = images / channel_max

1) np.arcsinh(images / 20.0)

• This is a non-linear compression, similar to logarithmic stretching.
• Astronomers use arcsinh because it:
  • Preserves faint structures (outer spiral arms, halos)
  • Prevents bright cores from blowing out white
• Empirically, arcsinh gave sharper features and improved ROI

2) channel_max = images.reshape(-1, 3).max(axis=0)

• The image tensor is flattened → shape (N, 3)
• We then compute the maximum intensity per channel (R/G/B separately)
• Result example: channel_max = [max_R, max_G, max_B]

3) images = images / channel_max

• Each image is normalized by its channel-wise maximum
• Ensures all channels are scaled to [0,1] range
• Prevents any color channel from dominating

Current Limitation

Even with tuning, anomaly ROC ≈ 0.51, meaning:

• Model struggles to separate normal vs. anomalous galaxies
• AE may be reconstructing anomalies too cleanly
• Low morphological contrast = hard boundary to detect

---

Next Step — Add AHUNT Pipeline

Found a method that directly targets galaxy anomaly detection:

Personalized anomaly detection using deep active learning (2023) — RASTI Journal
Could be integrated to significantly improve anomaly separation.

Recommended expansion: implement AHUNT downstream of VAE embeddings for more robust anomaly detection.