PCCP Workshop Series : Bayesian Deep Learning for Cosmology and Gravitational waves

4 mars 2020, 14:00 → 6 mars 2020, 17:00 Europe/Paris

Pierre-Gilles de Gennes Amphitheater (PCCP, APC laboratory, Université de Paris)

Machine learning attracts a lot of interest in the fields of cosmology and gravitational-wave astronomy and may potentially lead to major breakthroughs. Its adoption by the scientific community is increasing dramatically but it does not yet belong to the toolbox of 'off-the-shelf' algorithms. One of the reasons is that built-in uncertainty estimation, which is core to the evaluation of any scientific measurement and analysis, is not yet common in machine learning models.

Such limitation is on the verge to be overcome by the emergence of probabilistic machine learning models and algorithms. Among them, recent models called Bayesian neural networks, which combine machine learning and Bayesian statistics, use new (deep) neural networks architectures to enable Bayesian inference, and have received a great attention from the artificial intelligence community over the past few years.

This workshop will give the participants the opportunity to learn more about these emerging methods and how to use and exploit them in their research. The workshop program includes invited lectures and tutorials from major computer science experts and contributed talk and poster session aimed at sharing experience between physicists on the practical applications of machine learning.

 

The workshop is intended to researchers and students that are familiar with machine learning, use this type of algorithms for their own work, and want to learn about the advanced techniques related to Bayesian deep learning.

Workshop poster

mercredi 4 mars

14:00 → 14:30 Organisers: Welcome address George Smoot (Nobel prize 2006)

14:00 Logistics information

BNN_workshop_Logistics_precautions_intro.pdf

14:30 → 15:10 Invited talk

14:30 An Introduction to Bayesian Deep Learning

Bayesian Deep Learning (BDL) fills an important gap in the current deep neural networks, no matter powerful they are: in figurative terms, one could say BDL gives to AI the introspective ability to assess its own level of ignorance due to a lack of observations.
In more technical terms, BDL adopts the view of Bayesian statistics by replacing the weights of neural networks by distributions.
While this idea is nothing new, BDL has recently undergone new developments, thanks in particular to the seminal work of Y. Gal.
This presentation is designed to be a gentle introduction of the main concepts of BDL.
It introduces the required notions of Deep Learning and Bayesian statistics before developing the BDL framework.
It will not assume any particular piece of knowledge, but some general notions in machine Learning and Neural Networks.

Orateur: Frédéric Pennerath (CentraleSupélec)

BDL-introduction.pdf

15:10 → 15:50 Contribution talks

15:10 Star-galaxy separation via Gaussian Processes with Neural Network Dual Kernels

Orateur: Dr Imene Goumiri (llnl)

15:30 Detection of gravitational-wave signals from binary neutron star signals using machine learning

Orateur: Marlin Benedikt Schäfer (Albert Einstein Institut Hannover (AEI Hannover))

talk.pdf

15:50 → 16:10 Coffee break

16:10 → 16:50 Contribution talks

16:10 Deep learning dark matter map reconstructions and parameter inference with Dark Energy Survey data

Orateur: Dr Niall Jeffrey (École normale supérieure)

DeepMass_LFI.pdf

16:30 Neural networks estimation of the dense-matter equation of state from neutron-star observables

Orateur: M. Filip Morawski (Nicolaus Copernicus Astronomical Center of the Polish Academy of Sciences)

presentation.pdf

16:50 → 17:10 Lightning talks: 1 min presentation of posters

PhD or postdocs present quickly their work

17:10 → 17:50 Invited talk

17:10 Bayesian analysis and Supernova Photometric Cosmology

In the end of the 20th century type Ia supernovae provided the first evidence for accelerated cosmic expansion -- completely changing the cosmological model paradigm. Since then, the astronomical community has devote much of its resources to the construction of large scale sky surveys which are expected to achieve first light in the next few years. The upcoming Large Survey of Space and Time at the Vera Rubin Observatory (LSST) is one of the most ambitious of such experiments. In the new data paradigm raised by the next generation of surveys, which will deliver larger and more complex astronomical data than ever before, the methods of data analysis will need to be adapted to the new reality. In this talk, I will briefly describe the traditional pipeline for supernova cosmology, highlighting the new challenges to be faced and list a number of potential improvements already achieved by the application of Bayesian analysis. In particular, I will focus on how the combination of Bayesian techniques with adaptive machine learning algorithms can enable purely photometric supernova cosmology in the next decade.

Orateur: Dr Emille Ishida (LPC-UCA)

Ishida_BNN_workshop (2).pdf

18:30 → 19:30 Cocktail Reception

jeudi 5 mars

09:00 → 09:30 Coffee and pastries

09:30 → 11:00 Tutorial: Variational inference vs. MCMC

09:30 Tutorial: Bayesian deep learning - MCMC vs VI

Orateur: Dr Tom Charnock (IAP)

Bayesian deep learning tutorial

11:00 → 11:30 Coffee break and poster session

11:30 → 12:20 Round table: Variational inference vs. MCMC

Academia has a tradition of fostering interaction between scientific fields but not with companies. In the past years, we have witnessed a development of academic-type research teams within big tech companies, especially in machine learning. As machine learning is becoming a very hot topic in Physics (and quantum computing might also), we need these kind of interactions more than ever. How can we make those two worlds benefit from each other?

Présidents de session: Boris Leistedt (Imperial College), Edward Porter (APC/CNRS), François Lanusse (CEA/AIM), Junpeng Lao (Google), Tom Charnock (IAP)

12:20 → 14:00 Lunch on your own

14:00 → 14:40 Invited talk: Google / TensorFlow Probability

14:00 TensorFlow Probability

Orateur: Brian Patton (Google)

14:40 → 15:40 Contribution talks

14:40 Deep learning for a faster Hamiltonian Monte Carlo sampler

Orateur: M. Marc Arène (APC)

20200305_Bayesian_DL_workshop_MarcArene.pdf

15:00 Galaxy Zoo: Probabilistic Morphology through Bayesian CNNs and Active Learning

Orateur: M. Mike Walmsley (University of Oxford)

Paris BDL Conference Mike Walmsley Upload.pdf

15:20 Bayesian parameter estimation using conditional variational autoencoders for gravitational wave astronomy

Orateur: M. Hunter Gabbard (University of Glasgow)

Online version Paris_Bayesian_Deep_learning_workshop.pdf

15:40 → 16:10 Coffee break and poster session

16:10 → 17:10 Contribution talks

16:10 Denoising gravitational wave signals with a variational autoencoder

Orateur: Dr Philippe BACON (Laboratoire Astroparticule et Cosmologie)

denoisedseg.wav

noisyseg.wav

pbacon_presentation_apcworkshop.pdf

16:30 Solving source separation problem for LISA data analysis with Autoencoders

Orateur: Dr Natalia Korsakova (Observatoire de la Côte d'Azur)

16:50 The devil is in the details: interpreting probabilities from machine learning

Orateur: Alex Malz (German Centre of Cosmological Lensing)

AIMalz_BDL.pdf

17:10 → 18:00 Round table: Synergies between academia and tech companies: hardware, computing, algorithms and people

Academia has a tradition of fostering interaction between scientific fields but not with companies. In the past years, we have witnessed a development of academic-type research teams within big tech companies, especially in machine learning. As machine learning is becoming a very hot topic in Physics (and quantum computing might also), we need these kind of interactions more than ever. How can we make those two worlds benefit from each other?

Présidents de session: Alexandre Jean (Microsoft), Eric Aubourg (APC), Eric Chassande-Mottin (CNRS AstroParticule et Cosmologie), Laurent Daudet (Lighton), Marco Cuturi (Google), Volker Beckmann (CNRS / IN2P3)

vendredi 6 mars

09:30 → 12:30 Tutorial: TensorFlow Probability

09:30 TensorFlow Probability - Brian Patton, Junpeng Lao (TensorFlow Probability - Google)

Notebook on GW detection with sampling techniques

Notebook on PhotoZ estimation with Variational Inference

10:50 Coffee break

11:10 TensorFlow Probability - Brian Patton, Junpeng Lao (TensorFlow Probability - Google)

12:30 → 14:00 Buffet @ Microsoft

14:00 → 15:35 Invited talk: Graphcore, Microsoft Azure, LightOn

14:00 Graphcore: Innovation in Machine Intelligence Hardware & Software

Orateur: Alexander Titterton (Graphcore)

14:45 Large-scale Machine Learning on LightOn’s Optical Processing Units

Orateur: Laurent Daudet (Light On)

15:15 Microsoft Azure

Orateur: Alexandre Jean (Microsoft)

15:35 → 16:00 Organisers: Wrap-up

15:35 Organizers wrap-up

16:00 → 16:30 Coffee break and networking