Classifier Performance Stats

[a-desmons]

This issue lists the performance stats after adding a classifier layer to our SSML model.

Total number of objects in data set: 45,128
Number of objects with known tidal features: 187

• Images were first encoded and the Classifier Layer was trained for 500 epochs
• The Classifier outputs a single number between 0 and 1, with 0 signifying "object with no tidal features", and 1 signifying "object with tidal features"

From the 187 images with known tidal features, 12 were set aside for testing performance and testing for overfitting - These are shown here

All images with known tidal features (i.e. the 187) were found within the first 7038 images with a ranking closest to 1.

The table below shows:

• Test image number: The image used for testing
• Similarity index: Its ranking determined by the classifier, with 0 being the "best" and 45,128 being the "worst"

Test image number	Similarity Index
1	3982
8	307
18	4030
27	3800
76	451
86	997
97	418
117	21
119	7037
161	869
163	113
177	391

And a histogram of the classifier's performance in finding the images with known tidal features:

[a-desmons]

Comparing the performance of just the model (with single galaxy as query) to model+classifier.

x-axis: Similarity index in thousands (i.e. 1 = 1000)
y-axis: Fraction of galaxies with tidal features found

Zooming in on first half of the plot:

[a-desmons]

Sampling results by taking 6 images every 10,000 similarity-ranked images