Looking at a Black Hole with Python

Looking at a Black Hole with Python

A brief note on how deep learning and specially GAN's can be used to reconstruct the VLBI Images efficiently. And how Python can help in that!

View talk here: https://cloudproject.hosted.panopto.com/Panopto/Pages/Viewer.aspx?id=8b73aab5-428e-49ac-bff8-aa9b015eafea

Lightning talk delivered at Python in Astronomy 2019 @ Space Telescope Science Institute, Baltimore, United States of America

Be209241ca741792b3071d0dd5c55771?s=128

Shreyas Bapat

July 31, 2019
Tweet

Transcript

  1. Looking at a Black Hole with Python Shreyas Bapat (@astroshreyas)

    PyAstro 2019 @ STScI Baltimore (USA), 31 July 2019
  2. The BIG Problem Data! 1. (2nd Doesn’t exist)

  3. http://vlbiimaging.csail.mit.edu/

  4. The incomplete fourier space!

  5. Inverse Modeling ▪ Put 0’s in every other place ▪

    Take a simple Inverse Fourier Transform. ▪ Tada! Perform some extra steps to make the image less blurry!. 5
  6. Regularized Maximum Likelihood ▪ Predicting what values would come at

    the empty places. ▪ Basically one of the technique used GMMs! ▪ Inverse fourier transform. ▪ Tada! No need to perform any more steps!. 6
  7. 7

  8. Don’t bash me xD! But why not try with Deep

    Neural Networks! 8
  9. CNN’s Convolutional Neural Networks have outperformed almost every other technique

    in image-based problems! 9
  10. One night in 2014, Ian Goodfellow went drinking to celebrate

    with a fellow doctoral student who had just graduated. He had an idea: What if you pitted two neural networks against each other? Generative Adversarial Networks (GAN) were born! 10
  11. MR Imaging Very similar to VLBI. More number of points

    in Fourier Space, more the time taken for one image. My experience Results Using Residual Learning, Got really sharp images from undersampled fourier space. 11
  12. My process is easy Fill the remaining places with 0’s

    Manipulate the 0’s to some other value Take the Fourier Transform 12
  13. 13

  14. ~93% Reconstruction Accuracy! 14

  15. What if we use GANs? 15

  16. GAN : A minimax game between two neural networks Generator

    tries to generate the image! Discriminator tries to judge if the image is good to go! 16
  17. 17

  18. 18

  19. 19

  20. Stay in touch! GitHub : @shreyasbapat Twitter : @astroshreyas Email

    : hello@shreyasb.com 20
  21. 21