Deep Vacuum Cleaner

Deep Vacuum Cleaner

Deconvolving radio astronomy images using a convolution network.

978e79ad01185b39efcfca1482f0f819?s=128

Gijs Molenaar

August 02, 2018
Tweet

Transcript

  1. Deep Vacuum Cleaner Gijs Molenaar

  2. Deep Learning, why? • Deep learning seems to perform well

    on image to image tasks • Using convolution network for learning inverse of convolution • Can learn any function given enough nodes in hidden layer • Recursive networks are Turing complete • Interesting to get experience with
  3. The architecture • No idea! • Auto encoder? Recursive Neural

    Machines? LSTM?
  4. Pix2pix

  5. generative adversarial U- shape auto encoder network • Actually 2

    networks • Generator • Discriminator (used during training)
  6. The generator

  7. The discriminator

  8. Putting it together

  9. Training pipeline • Spiel • CWL based https://github.com/gijzelaerr/spiel/

  10. None
  11. Simulation Job

  12. None
  13. In short • Works on Linux, OS X (Containers) •

    Implicit parallelisation • Support for various schedulers (Slurm, Mesos) • Open Standard
  14. Training data • Deep learning usually needs a lot of

    data • Point sources only • 256x256 fits • 2000 Meerkat16 datasets • We can flip and rotate those! Boom, more data. • Randomised: sky position, synthesis time, channel width, fluxes
  15. Pre/post processing • Scale fluxes to 1 / -1 •

    Rescale output back
  16. None
  17. None
  18. None
  19. Test set RMS residuals

  20. What (doesn’t) works • All bright sources found • PSF

    independent • Fluxes are wrong • Network is good at generating images that ‘look like’ what we want
  21. Experiment PSF • Add PSF to network as channel •

    Network converges faster • Probably only helps in center of image
  22. Experiment Vis2img • Statistically good looking nonsense

  23. Experiment vis2vis • training… • Error based image space

  24. Time performance • Deconvolving deep2 image • 1089 tiles •

    30 seconds total • 10 seconds ‘warming up’ • 20 ms per tile • Exported model is about 100 megabytes
  25. Applications • Quick and dirty cleaning? • Quickly get statistical

    estimates?
  26. Future work • Better train data • Try different architecture

    • Infrastructure is there • Collaborate with Deep Learning experts • Recurrent Inference Machines for Solving Inverse function ill- posed problems • https://arxiv.org/abs/ 1706.04008
  27. Fin