Automated Lensing Learner (ALL): Identifying Strong Lenses · Not great model performance on 1 yr...
Transcript of Automated Lensing Learner (ALL): Identifying Strong Lenses · Not great model performance on 1 yr...
Automated Lensing Learner (ALL):
Identifying Strong Lenses
Camille Avestruz
Joint Fermi and KICP Fellow
Provost's Postdoctoral ScholarUniversity of Chicago
Cosmology on Safari February 14, 2017with Nan Li (Argonne/KICP) and Matthew Lightman (JP Morgan Chase)
Lensing Pipeline Mock Data Model Performance Conclusions
Galaxies are Strong Lenses
Source plane Image plane
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Galaxies and Structure Formation Tool
Magni�es high z galaxies
Mass reconstruction
Rau+14
DE probe
Meneghetti+04
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Galaxies and Structure Formation Tool
Magni�es high z galaxies
Mass reconstruction
Rau+14
DE probe
Meneghetti+04
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Galaxies and Structure Formation Tool
Magni�es high z galaxies
Mass reconstruction
Rau+14
DE probe
Meneghetti+04
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
ALL Pipeline: Extract Features and Train a Model
1
Edges
1
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
ALL Pipeline: Extract Features and Train a Model
1
Edges
1
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
ALL Pipeline: Extract Features and Train a Model
1
Edges
1
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
ALL Pipeline: Extract Features and Train a Model
1
Edges
1
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
ALL Pipeline: Extract Features and Train a Model
1
Edges
1
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
ALL Pipeline: Extract Features and Train a Model
1
Edges
1
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
ALL Pipeline: Extract Features and Train a Model
1
Edges
1
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Detect Edges via Histogram of Oriented Gradients
Compute the gradient vector for each pixel
~v = [38, 38]
θ = 45o
De�ne a cell of npix × npix (e.g. 8x8 pixels per cell)Bin cell's 64 pixel gradients
45o
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Detect Edges via Histogram of Oriented Gradients
Compute the gradient vector for each pixel
~v = [38, 38]
θ = 45o
De�ne a cell of npix × npix (e.g. 8x8 pixels per cell)Bin cell's 64 pixel gradients
45o
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Detect Edges via Histogram of Oriented Gradients
Compute the gradient vector for each pixel
~v = [38, 38]
θ = 45o
De�ne a cell of npix × npix (e.g. 8x8 pixels per cell)Bin cell's 64 pixel gradients
45o
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Detect Edges via Histogram of Oriented Gradients
Compute the gradient vector for each pixel
~v = [38, 38]
θ = 45o
De�ne a cell of npix × npix (e.g. 8x8 pixels per cell)
Bin cell's 64 pixel gradients
45o
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Detect Edges via Histogram of Oriented Gradients
Compute the gradient vector for each pixel
~v = [38, 38]
θ = 45o
De�ne a cell of npix × npix (e.g. 8x8 pixels per cell)
Bin cell's 64 pixel gradients
45o
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Detect Edges via Histogram of Oriented Gradients
Compute the gradient vector for each pixel
~v = [38, 38]
θ = 45o
De�ne a cell of npix × npix (e.g. 8x8 pixels per cell)
Bin cell's 64 pixel gradients
45o
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Logistic Regression to Train a Model Classi�er
Linear classi�er to draw a hyperplane boundary infeature space (2D example)
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Space based tests with mock HST data
PICS: Li et al., 2016
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Ground based tests with mock LSST data (10 yr)
LensPOP: Colletti et al., 2013
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Ground based tests with mock LSST data (1 yr)
LensPOP: Colletti et al., 2013
10-year epoch Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Use Trained Model to Rank Order the Test Sample
High score:.999
Mid score:.402
Low score:.05
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Use Trained Model to Rank Order the Test Sample
High score:.999
Mid score:.402
Low score:.05
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Use Trained Model to Rank Order the Test Sample
High score:.999
Mid score:.402
Low score:.05
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Assess model performance with Receiver Operating
Characteristic (ROC) curve
False positive rate
Truepositiverate
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Set threshold based on the ROC curve
False positive rate
Truepositiverate
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Great model performance on mock HST data
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Good model performance on 10 yr LSST data
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Not great model performance on 1 yr LSST
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Highly magni�ed images have high scores
0 5 10 15 20
magnitude
0.0
0.2
0.4
0.6
0.8
1.0
score
150.6
170.5
190.4
210.3
230.2
250.1
270.0
289.9
309.8
329.7
velo
city
dis
pers
ion
Image Magni�cation
Avestruz ALL - Lens�nding
Lensing Pipeline Mock Data Model Performance Conclusions
Conclusions
Machine learning techniques will be necessary in
forthcoming datasets
ALL is a generalizable lensing image classi�er pipeline
Our HOG + Linear Regression is a fast and easy tool to
classify galaxy scale strong lensing systems
Avestruz ALL - Lens�nding