Autoencoder-Based Galaxy Image Search: CHUNGNAM NATIONAL UNIVERSITY

A New Approach to Morphological DEPARTMENT OF ASTRONOMY AND SPACE SCIENCE & RESEARCH
INSTITUTE OF NATURAL SCIENCES

Classification and Galaxy Research Methods SUK KIM

2024. 1. 25.
AutoEncoder:
A Gateway to Efficient
Data Representation

Autoencoders: Specialized feedforward

neural networks
Design: Crafted for input reconstruction
Objective: Input replication through output
loss

epoch
 “Code” or “latent space” or “latent features”

0.7 0.6 0.1 0.2 0.8 0.9 0.3

… 0.2 0.4 0.1 0.1 0.2 0.2 0.6

Number of images
0.3 0.2 0.1 0.5 0.8 0.3 0.1
… 0.4 0.1 0.1 0.5 0.2 0.9 0.3
0.7 0.5 0.2 0.1 0.2 0.3 0.1
… 0.9 0.4 0.2 0.1 0.1 0.1 0.6
0.7 0.6 0.1 0.2 0.8 0.9 0.3
… 0.2 0.4 0.1 0.1 0.2 0.2 0.6

…
0.7 0.5 0.2 0.1 0.2 0.3 0.1
… 0.9 0.4 0.2 0.1 0.1 0.1 0.6
0.7 0.6 0.1 0.2 0.8 0.9 0.3
… 0.2 0.4 0.1 0.1 0.2 0.2 0.6
Autoencoders: Pioneers in Data Dimensionality Reduction and
Accurate Reconstruction

Efficient Compression: Autoencoders distill essential information, encoding inputs

into a more compact, lower-dimensional form.
Reconstruction: They then use this encoded essence to recreate the original input as
accurately as possible.
What are the potential applications of an
autoencoder model?
1. Search for similar images
- Tool for Searching a similar Galaxy Image based on a Convolutional Autoencoder using Similarity
(TSGICAS; see Seo et al. 2024)

2. Morphological classification
- Seo et al. 2024 (Semi-supervised learning)
- Clustering

3. Denoising

4. Analyze the changes in an image as a function of wavelength

 Tool for Searching a similar Galaxy Image based on a
Convolutional Autoencoder using Similarity
(TSGICAS; see Seo et al. 2024)
DATABASE
“Code” or “latent space” or “latent features”
Target galaxy
0.7 0.6 0.1 0.2 0.8 0.9 0.3
… 0.2 0.4 0.1 0.1 0.2 0.2 0.6

Number of images
0.3 0.2 0.1 0.5 0.8 0.3 0.1
… 0.4 0.1 0.1 0.5 0.2 0.9 0.3
0.7 0.5 0.2 0.1 0.2 0.3 0.1
… 0.9 0.4 0.2 0.1 0.1 0.1 0.6
Di
0.7 0.6 0.1 0.2 0.8 0.9 0.3
… 0.2 0.4 0.1 0.1 0.2 0.2 0.6

…
0.7 0.5 0.2 0.1 0.2 0.3 0.1
… 0.9 0.4 0.2 0.1 0.1 0.1 0.6
0.7 0.6 0.1 0.2 0.8 0.9 0.3
… 0.2 0.4 0.1 0.1 0.2 0.2 0.6

latent features of target galaxy

Background 0.7 0.6 0.1 0.2 0.8 0.9 0.3

… 0.2 0.4 0.1 0.1 0.2 0.2 0.6 Ti
image of the
target latent features of background image

0.7 0.6 0.1 0.2 0.8 0.9 0.3

… 0.2 0.4 0.1 0.1 0.2 0.2 0.6
Bi
Tool for Searching a similar Galaxy Image based on a
Convolutional Autoencoder using Similarity
(TSGICAS; see Seo et al. 2024)
Tool for Searching a similar Galaxy Image
based on a Convolutional Autoencoder
Project of the Observation Survey
using Similarity (TSGICAS)

Target galaxy

TSGICAS
Simulation
Morphological classification
Morphological classification
SBc
Morphological classification
- clustering
 Deciphering Morphological Characteristics of sub-halo Dark
Matter in Simulations via Autoencoder Analysis

IllustrisTNG
project
 K-mean clustering

PC6

“Code” or “latent space” or “latent features”

0.7 0.6 0.1 0.2 0.8 0.9 0.3

… 0.2 0.4 0.1 0.1 0.2 0.2 0.6

Number of images
0.3 0.2 0.1 0.5 0.8 0.3 0.1
… 0.4 0.1 0.1 0.5 0.2 0.9 0.3

PC5 0.7 0.5 0.2 0.1 0.2 0.3 0.1

… 0.9 0.4 0.2 0.1 0.1 0.1 0.6
0.7 0.6 0.1 0.2 0.8 0.9 0.3
… 0.2 0.4 0.1 0.1 0.2 0.2 0.6

…
0.7 0.5 0.2 0.1 0.2 0.3 0.1
… 0.9 0.4 0.2 0.1 0.1 0.1 0.6
0.7 0.6 0.1 0.2 0.8 0.9 0.3
… 0.2 0.4 0.1 0.1 0.2 0.2 0.6

PC1
 1 A
Group 5

1
6

2 6

2
3 7
Group C
Group B 7
3
4

4
5
Deciphering Morphological Characteristics of sub-halo Dark Matter in Simulations via Autoencoder Analysis

Group A Group C Group B

Denoising Autoencoder

Add noise

loss
Exposer time : 68 h Exposer time 40 m JCMT 850um image Low S/N Denoising High S/N
High S/N Low S/N

Denoising Image
Denoising 1D Spectra
I-GALFA data

Add noise
Analyze the changes in an image as a function of
wavelength

Suppose the differences in image characteristics of the same

galaxy that change depending on the wavelength can be quantified.
In that case, it can be a suitable new parameter representing the
galaxy’s state.
Analyze the changes in an image as a function of
wavelength

0.7
0.70.6
0.60.1
0.1 0.2 0.80.9
0.20.8 0.90.3 …
…
…
0.2
0.20.4
0.40.1
0.1 0.1
0.1 0.2
0.20.2
0.20.6
…
0.3 0.6
…
0.7 0.6 0.1 0.2 0.8 0.9 0.3 0.2 0.4 0.1 0.1 0.2 0.2 0.6
0.80.9
0.90.3 0.2
0.20.4
0.40.1
0.1 0.1
0.1 0.2
0.20.2
0.20.6
…
0.7
0.70.6
0.60.1
0.1 0.2
0.20.8 0.3 0.6
0.7
0.70.6
0.70.6
0.1
0.60.1
0.2
0.1 0.2
0.8 0.9
0.80.9
0.20.8 ……
0.3
0.90.3
…
0.2
0.20.4
0.20.4
0.1
0.40.1
0.1
0.1 0.1
0.2
0.1 0.2
0.2
0.20.2
0.6
0.20.6
…
0.3 0.6
…
0.7 0.6 0.1 0.2 0.8 0.9
0.90.3
0.3 0.2
0.20.4 0.1 0.1 0.2 0.2 0.6
0.2 0.4 0.1 0.1 0.2 0.20.6
0.4 0.1 0.1 0.2 0.2
…
0.7 0.6 0.1 0.2 0.8
0.7 0.6 0.1 0.2 0.8 0.9 0.3 0.6
0.7 0.6 0.1 0.2 0.8 0.9 0.3 0.2 0.4 0.1 0.1 0.2 0.2 0.6

latent features
Wavelength
Machine Learning for Enhanced Photometric
Redshift Estimation
Determining Member Galaxies in Clusters and Groups
Using Photometric Redshifts in the Local Universe
Objective: To achieve a photometric redshift error of 0.001 (equivalent to 300 km/s) for
galaxies with z < 0.1 in the local universe.

Purpose: This study aims to accurately determine member galaxies of clusters and groups in
the local universe using photometric redshift measurements.
 Deep Learning Test
#40 medium-band filter
#20 medium-band filter zCOSMOS catalog / matched spac-z

self.layer_1 = nn.Linear(num_features, 190) self.layer_1 = nn.Linear(num_features, 100)

self.layer_2 = nn.Linear(190, 100) self.layer_2 = nn.Linear(100, 200)
self.layer_3 = nn.Linear(100, 50) self.layer_3 = nn.Linear(200, 400)
self.layer_4 = nn.Linear(50,30) self.layer_4 = nn.Linear(400,800)
self.layer_out = nn.Linear(30, 1) self.layer_5 = nn.Linear(800,400)
self.relu = nn.ReLU() self.layer_6 = nn.Linear(400,200)
self.layer_7 = nn.Linear(200,100)
self.layer_out = nn.Linear(100, 1)

self.relu = nn.ReLU()
Deep Learning Test
Resnet50 (CNN)
#40 medium-band filter
zCOSMOS catalog
"Using direct image data in machine
learning significantly improves the
accuracy of photometric redshift
estimation."
Key Proposal Highlights:
Machine Learning Integration:
Employing machine learning techniques to analyze and interpret the nuanced features in multi-wavelength galaxy images.
Systematic Parameter Development:
Creating parameters to track and quantify the variations in galaxy characteristics across different wavelengths.
Galaxy Evolution Study:
Proposing a new research direction based on these systematically derived parameters.

Enhanced Photometric Redshift Measurement:

Utilizing a combination of photometry data and multi-wavelength imaging to significantly improve the accuracy of photometric
redshift estimations through machine learning.