Pattern recognition in multi wavelength astronomical data in the context of the SKA and its pathfinders
Tuesday, 19th of January 2021
I. Participants:
- Benoit Semelin
- Philippe Zarka
- Françoise Combes
- Martin Hardcastle
- Kevin Schmidt
- Francesco De Gasperin
- Rafaël Mostert
- Beatriz Mingo
- Lara Alegre
- Judith Croston
- Philippe Salomé
- Philip Best
- Kushatha Ntwaetsile
- Joshua Albert
- Erik Osinga
- Cyril Tasse
- David Cornu
- Baptiste Cecconi
- Laurent Lamy
- Jurjen de Jong
- Bonny Barkus
- Alan Loh
- Ulrich A. Mbou Sob
- Huub Rottgering
- Stéphane Corbel
- …
II. Program:
(CET times)14:00-14:10:
Welcome / Round table
14:10-15:30 Talks
15:40-17:00 Discussion
III. Talks: [7 minutes talk, 3 minutes questions]
- David Cornu: 3D extinction mapping of the Milky Way using CNNs
- Philippe Salomé : Yafits – A remote radio-data viewer – Demo (if people interested)
- Bonny Barkus: automated host identification and classification of extended radio sources – Part I: Intro and Ridgelines
- Beatriz Mingo: automated host identification and classification of extended radio sources – Part II: Morphological Classification + Conclusions
- Lara Alegre: Classification of LOFAR radio sources with machine learning.
- Rafaël Mostert: Radio component association using r-CNNs and optical cross-identification using CNNs.
- Kevin Schmidt: Deep Learning Approaches to Improve Imaging
- Kushatha Ntwaetsile: Rapid sorting of radio galaxy morphology using Haralick Features
- Sydil Kupa : Automating Image segmentation using CNN for MeerKAT Radio Galaxy Images
IV. Discussion
1. Machine Learning approaches:
1.1 Radio source morphology
- In the context of morphological classification: Are there rare types of objects and what is the dilution factor with frequent types of objects in our datasets ? Possible discussion on how to properly evaluate the effect of such objects in terms of prediction purity (proportion of false positives).
- How to handle morphological degeneracy (when there is some) ?
1.2 Multiwavelength morphology
- What would be the most efficient ML approaches for hyperspectral surveys ?
- Would it be possible to extract characteristic morphological features of a given class of object using generative models ? (e.g GAN)
1.3 Machine learning with Lofar Galaxy Zoo
- Given the data that we have from the first year of Lofar Galaxy Zoo, can we better select which sources should not go into the Zoo?
- Related: Can we use the LGZ data to train a ML algorithm to do the task, and then in cases where it is uncertain we put those subjects into the project.
2. Applications:
2.1 LoTSS:
- Where do we stand? Where do we go?
2.2 SKA challenges :
- CNN architectures for object detection
- Data pre-processing & augmentation
- Transfer learning from other radioastronomical surveys
3. Interpretation of the results?
- How to avoid falling into the usual traps ( e.g. resolution/sensitivity limits in different surveys yielding different results for the same source, overinterpretation of the results, interpretation within schemes that are now obsolete…)?
- How can we make sure that we don’t cause more problems than we fix?
- Using joint approaches for different surveys (e.g. SPARCS)?
4. Research program / Potential collaborations?
- Sharing astro dedicated ML methods tuning (e.g. dedicated ANN architectures modification sharing)
- Exploration of transfer learning between applications and surveys (Existing githubs?, test datasets?)