Towards probabilistic models for capturing uncertainty

Alexandre Boucaud

CNRS/IN2P3 - France

alexandre.boucaud@apc.in2p3.fr

Instrumental context

Euclid ESA satellite

15 000 sq. deg.

higher resolution than ground telescopes

3 instruments – visible + near-IR imaging

launched end of 2022

2

Current surveys

credit : SDSS

Upcoming surveys

credit : NGVS

Galaxy blending

z=0.1

z=0.2

courtesy H. Bretonnière

galaxies are "transparent" �=> no obscuration

measuring flux and shape when galaxies overlap is tricky

in our case a pixel can refer to �several objects

Galaxy blending

Galaxy blends emulated with real galaxy images

https://github.com/aboucaud/candels-blender

Blended galaxy segmentation

https://arxiv.org/abs/1905.01324

Galaxy segmentation with UNet

9

input images

(test set)

true

segmentation

predicted

segmentation

https://arxiv.org/abs/1905.01324

Flux estimation of blended galaxies

Using a classic convnet, directly on the blend galaxy images

performance is much better than traditional astro detection algorithms

https://arxiv.org/abs/1905.01324

Could we go from a fully deterministic network..

..to a probabilistic one ?

Thesis of Hubert Bretonnière

"Develop and implement deep learning-based image �processing algorithms for the morphology of galaxies�Euclid satellite"

co-supervision astronomer – software engineer

started last october

hubert.bretonniere@ias.u-psud.fr

​

https://github.com/SimonKohl/probabilistic_unet

https://arxiv.org/abs/1806.05034

https://arxiv.org/abs/1806.05034

https://arxiv.org/abs/1806.05034

https://arxiv.org/abs/1806.05034

TensorFlow Probability

TensorFlow Probability: Learning with confidence (TF Dev Summit '19)

https://www.tensorflow.org/probability

Probabilistic segmentation

aim at predicting a probability of blending between 2+ galaxies

can be applied to large images

ability to propose an absence of overlap

uses TensorFlow Probability

courtesy H. Bretonnière

input images

true

segmentation

predicted

segmentations

courtesy H. Bretonnière

• deterministic networks tend to hide the model uncertainty
• modifying your models to output probability distributions �is quite straightforward
• TensorFlow => TensorFlow Probability at least
• other tools exist (Pyro w/ PyTorch, PyMC3, etc..)
• such step might be necessary in order to use your ML model to perform Bayesian inference

Conclusions

​

ANR – “AstroDeep”

Recently got funding for the next 4 years

3 postdocs

1 PhD student

travel and computing

23

Astro experts

weak lensing

signal processing

image processing pipelines

Computer scientists

machine learning

neural networks

Markov models, random processes, bayesian networks…

• 3-day workshop including �talks, round tables and hands-on
• focus on Bayesian inference w/ NN
• advanced tutorials given by �TFProbability developers
• deadline for application: feb 2

Workshop in march

https://indico.in2p3.fr/event/19458

• 3-day course with tutorials in the afternoon (50%)
• first 2 days focused on traditional machine learning �(terminology, main algorithms, model comparison, etc.)
• last day focused on neural networks and deep learning

​

Registration starting on Feb 15, course last week of May in Villejuif (Paris)

Formation CNRS

with Sylvain Caillou (LIMSI)