CosmosClub: Sebastian Rojas Gonzalez (12/07/19)

Date: July 12th 2019, 11am

Speaker: Sebastian Rojas Gonzales (KU Leuven)

Title: Gaussian processes for simulation optimization

Room: Kepler


Abstract

The use of kriging metamodels (also known as gaussian processes) in simulation optimization has become increasingly popular during recent years. The majority of the algorithms so far uses the ordinary (deterministic) kriging approach for constructing the metamodel, assuming that observations have been sampled with infinite precision. This is a major issue when the simulation problem is stochastic: ignoring the noise in the outcomes may lead to inaccurate predictions. In this work, we propose a stochastic multiobjective simulation optimization algorithm that contains two crucial elements: the search phase implements a kriging method that is able to account for the inherent noise in the outputs when constructing the metamodel, and in the identification phase uses a Bayesian multiobjective ranking and selection procedure in view of maximizing the probability of selecting the true non-dominated points by optimally allocating the available computational budget. We evaluate the impact of these elements on the search and identification effectiveness on a set of artificial test problems with varying levels of heteroscedastic noise. Our results show that the characterization of the noise is crucial in improving the prediction efficiency; yet, the allocation procedure appears to lose effectiveness in settings with high noise. This emphasizes the need for further research on multiobjective ranking and selection methods.

CosmosClub: Doogesh Kodi Ramanah (20/06/2019)

Date: June 20th 2019, 11am

Speaker: Doogesh Kodi Ramanah (IAP)

Title: Fast complex dynamics emulators for cosmological inference

Room: Cassini


Abstract

I will present an overview of our recent work in developing various aspects of Bayesian forward modelling machinery for an optimal exploitation of state-of-the-art galaxy redshift surveys. I will focus on the development of a generative model for mapping dark matter simulations to 3D halo fields using physically motivated neural networks. We employ the Wasserstein distance as a metric to train our halo painting emulator and demonstrate its efficacy in predicting 3D halo distributions using summary statistics such as the power spectrum and bispectrum. I will subsequently briefly review our novel cosmological parameter inference framework that extracts several orders of magnitude more information from the cosmic expansion relative to standard approaches, and a sophisticated likelihood that is robust to unknown foreground contaminations.

CosmosClub: Isabella Carucci (13/06/2019)

Date: June 13th 2019, 10:30am (/!\ unusual time)

Speaker: Isabella Carucci (CosmoStat)

Title: 21cm intensity mapping: using cosmic neutral hydrogen as tracer of large scale structure

Room: Kepler


Abstract

This is an opportunity to introduce my past work to the group. I’ll start by briefly illustrating what we mean by intensity mapping (IM) and why to use the 21cm radiation coming from neutral hydrogen. I’ll then move to what its observational status is and how we can model this signal. In particular, I’ll show how IM will be a remarkable test both for dark energy and dark matter models, presenting forecasts for the bounds that the SKA telescope will be able to uniquely set. On the other hand, this signal is literally buried under foregrounds (about 4 orders of magnitude more intense). In this spirit, I’ll conclude by sketching what our plans are to use and optimise Cosmostat-developed tools for tackling the foreground and instrumental systematics problems of IM.

CosmosClub: Jean-Baptise Bayle (23/05/2019)

Date: May 23rd 2019, 11am

Speaker: Jean-Baptise Bayle (APC)

Title: Space-based Detection of Gravitational Waves with LISA: Opening the Low-Frequency Part of the Gravitational Spectrum [slides]

Room: Cassini


Abstract

Einstein’s theory of General Relativity offers a powerful picture of the world, where the gravitational force is only the consequence of the curvature of spacetime. Gravitational waves are very tiny ripples in the fabric of spacetime, propagating at the speed of light. They do not interact much with the matter encountered on their way, making them very good messengers to probe astrophysical events at a distance... but also very difficult to detect. Scientists have looked for gravitational waves on Earth for more than 40 years and have developed incredibly precise instruments. In 2015, the first direct detection of such waves was announced by the LIGO and Virgo community: gravitational astronomy had opened a new observational window on the distant Universe. Unfortunately the limited sensitivities of such ground-based detectors does not enable the observation of the most massive object.

In two decades, the Laser Interferometer Space Antenna (LISA) mission will join the network of ground detectors. LISA aims to measure gravitational waves in the millihertz range, to help answer numerous astrophysical, cosmological and theoretical questions. Three spacecraft will orbit the Sun in a nearly equilateral triangular formation, with arm lengths of about 2.5 million kilometers. Each spacecraft will host two spacetime probes. Laser beams, exchanged between the spacecraft, will be reflected upon the probes and then made to interfere, in order to detect passing gravitational waves.

During this conversation, I will briefly recap how gravitational waves arise from Einstein equations, and present the astrophysical and cosmological sources expected throughout the gravitational spectrum. I will then describe the main detection principles, from ground-based interferometry to pulsar timing. I will then focus on the LISA mission, and discuss my PhD work on data pre-processing and instrumental modeling.

 

CosmosClub: Niall Jeffrey (25/04/2019)

Date: April 25th 2019, 11am

Speaker: Nial Jeffrey (UCL, currently visiting CosmoStat)

Title: Deep Learning dark matter maps from Dark Energy Survey (DES) weak lensing data

Room: Cassini


Abstract

Reconstructed density fields from weak lensing are rich in information about cosmological parameters and models of the Universe — including a large non-Gaussian component that cannot be accessed using traditional 2-point statistics. I will present a new method based on Deep Learning to reconstruct dark matter maps from weak lensing data with higher accuracy. Weak lensing map reconstruction is “ill-posed”, troubled by survey masks and galaxy “shape noise”. With DES SV data we showed that by implementing physically-motivated priors (Gaussian field or halo model), substantial improvements are made over standard approaches. Such advanced methods are still limited due to their prior distributions; non-linear density fields have no simple closed form that can be used as a prior. Deep Learning methods can directly learn the underlying structure of the signal, noise and mask from realistic simulations. By combining Deep Learning methods with a physically motivated closed-form prior, improved reconstruction is guaranteed.

Journal Club#5: Image reconstruction, H0 measurements and animated objects

CosmosClub: Antoine Labatie (11/04/2019)

Date: April 11th 2019, 11am

Speaker: Antoine Labatie

Title: Characterizing Well-behaved vs. Pathological Deep Neural Networks [paper]

Room: Kepler


Abstract

We introduce a novel approach, requiring only mild assumptions, for the characterization of deep neural networks at initialization. Our approach applies both to fully-connected and convolutional networks and easily incorporates the commonly used techniques of batch normalization and skip-connections. Our key insight is to consider the evolution with depth of statistical moments of signal and noise, thereby characterizing the presence or the absence of pathologies in the hypothesis space encoded by the choice of hyperparameters. We establish: (i) for feedforward networks with and without batch normalization, depth multiplicativity inevitably leads to ill-behaved moments and pathologies; (ii) for residual networks with batch normalization, on the other hand, skip-connections induce power-law rather than exponential behaviour, leading to well-behaved moments and no pathology.