Algorithmic VLBI Baseline
Selection Sasha Safonova
Victor Pankratius, Michael Gowanlock, Justin Li MIT Haystack Observatory REU
11 August 2016
Image credit: ESO/B. Tafreshi http://www.eso.org/public/images/potw1253a/
What are VLBI baselines?
Goal
“The dishes of the SKA will produce
10 times the global internet traffic.”1
1 https://www.skatelescope.org/amazingfacts/ August 9, 2016 Image credit: SPDO/Swinburne Astronomy Productions
Question: How to select key baselines?
Preserve key observable information
Decrease processing time
Image credit: SPDO/Swinburne Astronomy Productions
Goals
Tools
Genetic Algorithm
(GA)
Structural Similarity (SSIM)
Masking
Machine Learning
Guidance by the Researcher Evaluation
Case studies
Optimization Approach
Application
Case studies
Horsehead Nebula:-Dense with a lot of structure -Challenging current image
reconstruction methods -Inspires masking
Simulated point source:Sparse
(Most image reconstruction techniques focus on sparse images)
Case studies allow us to assess how our method’s efficacy varies by object type.
Image source: astropy.org
Masks let the researcher emphasize important features
Genetic Algorithm
GA trains on the masked image with the goal of retaining an image’s salient features.
Intelligent Baseline Selection for Radio Interferometric Imaging
Genetic Algorithm
All baselines
Selected baselines
Create a population of uv baselines
Randomly pick first generation
Select the fittest individuals
Perform crossover and mutation
Form next generation
Baseline coverage
Simplified baseline coverage
Pick the first generation’s uv coverage randomly
Randomly pick first generation
Select the fittest individuals
Perform crossover and mutation
Form next generation
SSIM=0.7
SSIM=0.1
SSIM=0.8
SSIM=0.9Randomly pick first generation
Select the fittest individuals
Perform crossover and mutation
Form next generation
Find the best individuals using structural similarity (SSIM)
SSIM=0.7
SSIM=0.1
SSIM=0.8
SSIM=0.9Randomly pick first generation
Select the fittest individuals
Perform crossover and mutation
Form next generation
Find the best individuals using structural similarity (SSIM)
The best individuals mate via crossover
SSIM=0.9
SSIM=0.8
Randomly pick first generation
Select the fittest individuals
Perform crossover and mutation
Form next generation
The best individuals mate via crossover
SSIM=0.9
SSIM=0.8
Randomly pick first generation
Select the fittest individuals
Perform crossover and mutation
Form next generation
The best individuals mate via crossover
Randomly pick first generation
Select the fittest individuals
Perform crossover and mutation
Form next generation
SSIM=0.9
SSIM=0.8
Create new, fitter generations until we reach goal SSIM
Randomly pick first generation
Select the fittest individuals
Perform crossover and mutation
Form next generation
Initial results
Our approach generally
outperforms random baseline selection on
the synthetic disk image
Evaluation on synthetic disk image
Training image
Our approach generally outperforms
random baseline selection on the
Horsehead Nebula
Evaluation on the Horsehead Nebula
Suppose one observing run requires 1800 terabytes of
storage.
Image author: Scott Schiller
How much information can we retain with selected optimized baselines?
How much information can we retain with selected optimized baselines?
How much information can we retain with selected optimized baselines?
How much information can we retain with selected optimized baselines?
How much information can we retain with selected optimized baselines?
How much information can we retain with selected optimized baselines?
How much information can we retain with selected optimized baselines?
How much information can we retain with selected optimized baselines?
How much information can we retain with selected optimized baselines?
Conclusion
• With increasing number of baselines, data volumes make processing more challenging
• In certain experimental scenarios, scientists might choose to operate with fewer baselines (e.g., for quick previews)
• In some cases similar image quality can be obtained with fewer baselines
• Our exploration shows promising results, optimized selection performs better than random selection
• Applicability varies by type of observable object
• More case studies are forthcoming: • Supernova 1993J • Solar dynamics • M31
Conclusions
Image credit: NASA, ESA, and G. Bacon (STScI)
Acknowledgements
Michael GowanlockJustin Li
Victor Pankratius
Phil Erickson Vincent Fish Frank Lind
Big data radio science group