In this project, I want to find a unsupervised way to seperate galaxies and stars. TODOs:

VAE + HypCol vs. CNNs + surrogate class + HypCol

https://github.com/EdwardJKim/unsupervised-dl4astro/blob/master/notebooks/kmeans.ipynb

on galaxy zoo dataset

on SDSS dataset

to cut figures of galaxy zoo dataset

use manifold learning method to imporve the accuracy of classification result. use both 2D and 3D hidden variables as inputs

use several different VAE structures to help clustering/hidden layer classification process

Kmeans clustering with the original normalization method(-mean,/max)

Report for week 4

some of the code is credit to Edward Kim:

https://github.com/EdwardJKim/unsupervised-dl4astro/

Compare size/luminosity of objects and the hidden variables

seems only one hidden variable is useful

can not use such hidden variables to predict the redshift

Model saved in Classification+result+is+better+when+I+use+5+channels.ipynb

A good Classification result with ROC curve

The overall acc reached 88.8%

together with the saved parameters: Correct_normalization_great_segmentation0807.h5.zip

Only use the 4th channel (i channel) instead use all of the 5 channels in segmentation. Got much better result than before

Reached the highest accuracy of 88.8%

Combined the results by Aug.8

Better segmentation result when I use AE instead of VAE.(It should be like this)

and some result are saved in OneNote

The result here is pretty like when I used L2 regularizer

Compared several reasonable loss functions.

I calculate those loss function using Mathematica. The .nb files are also uploaded