Abstracts
Monday 10th June
Thomas Hofmann (ETH Zürich, Institute for Machine Learning)
Title:
Advances in Deep Learning Models for Cosmology
Abstract:
In recent years, deep neural networks have been used successfully in the field of cosmology. The most popular class of models are convolutional neural networks (CNNs), which have been applied to prediction tasks from 2D or 3D fields of simulated or observed data, e.g. in order to constrain cosmological parameters from convergence maps. Another class of models that has shown promise are generative or conditional generative models such as generative adversarial networks (GANs), which have been employed with the aim to bypass or augment expensive simulations.
The talk will provide an overview of more advanced modeling options for both CNNs and GANs, along with a discussion of their pros and cons. The aim is to encourage researchers in cosmology to better adapt models to their needs and data by considering more advanced modeling options.
Michelle Ntampaka (Harvard CFA)
Title:
The Role of Machine Learning in the Next Decade of Cosmology
Abstract:
Machine learning permeates our daily lives -- performing tasks from identifying the people in a photograph to suggesting the next big purchase -- but how will it change the way we do research? The last decade has seen a remarkable rise in interdisciplinary machine learning (ML)-based cosmology research. I will overview some of the ways in which ML offers enticing improvements in the ways we can interpret data, with a focus on methods to understand and interpret observations of large scale structure. The next decade will see a continued rise in data-driven discovery as methods improve and data volumes grow, and I will discuss opportunities and challenges for realizing the full potential of ML in cosmology.
Zoltan Haiman (Columbia University)
Title:
Finding and using beyond-Gaussian information in weak lensing
Abstract:
Weak lensing (WL) is a promising method to study the properties of dark energy and dark matter, test departures from general relativity, and measure the total mass of neutrinos. The first large WL experiments have recently been performed both in the context of galaxy lensing and the lensing of the cosmic microwave background (CMB). These experiments have provided independent evidence for dark energy, and offered valuable cross-checks between galaxy and CMB lensing measurements. However, WL probes the projected dark matter structures which, on small scales, are highly non-linear and non-Gaussian. This presents an exciting yet challenging task: how can one extract the maximum amount of information from lensing observables, while minimizing the impact of systematic errors? This will be especially important for forthcoming large WL datasets with over a billion galaxies, delivering unprecedented precision. I will review progress to date on this question, including the use of various non-Gaussian statistics. I will also introduce promising recent work utilizing deep learning, which has shown the potential to outperform human-designed statistics, possibly quite significantly.
Jose Manuel Zorrilla Matilla (Columbia University)
Title:
Using ML to extract non-Gaussian information from weak lensing datasets
Abstract:
Upcoming weak lensing (WL) surveys will make available unprecedented amounts of data, raising the need to develop methods to extract optimally all the cosmological information encoded in them. WL datasets contain significant non-Gaussian information, as has been shown through the use of observables beyond second order statistics, such as higher-order correlation functions, Minkowski functionals or lensing peaks. It is not clear, though, what’s the ceiling on the information that can be extracted. Neural networks have been proven highly effective as a means to extract information from highly diverse datasets. I will share our experience applying convolutional neural networks to simulated WL data, how they allow to by-pass the need to design of specific statistics while extracting more information than some non-Gaussian statistics, and some challenges for the future.
Balint Armin Pataki (Eötvös Loránd University)
Title:
Estimation of cosmological parameters from weak lensing simulations with CNNs
Abstract:
According to our current understanding, a significant fraction of the Universe is made up of dark matter. Since it interacts only via gravitation it is not possible to directly observe dark matter. Gravitating matter (both baryonic and the dark matter) slightly distorts the apparent shapes of background galaxies via weak lensing. Due to the nonlinearities on small scales, traditional cosmological statistics cannot extract all the possible information concerning the distribution and structure of the gravitating matter. In the early 2010s convolutional neural networks (CNN) revolutionized almost all the image recognition related tasks by heavily outperforming previous attempts on various fields. Here we present a convolutional neural network that gives significantly better estimates of the (σ8, Ωm) cosmological parameters from simulated weak lensing convergence maps generated by full N-body simulations and ray-tracing, than state-of-art methods. We also show that the CNN is able to yield significantly stricter constraints of (σ8, Ωm) cosmological parameters than the power spectrum at angular scales and shape noise levels relevant for future observations. Furthermore, we demonstrate that at shape noise levels achievable in future space surveys, the CNN yields 1.4-2.1 times smaller contours than peak count methods.
Janis Fluri (ETH Zürich, IPA)
Title:
Cosmological constraints with deep learning from KiDS-450 weak lensing maps
Abstract:
Convolutional Neural Networks (CNN) have recently been demonstrated on synthetic data to improve upon the precision of cosmological inference. In particular they have the potential to yield more precise cosmological constraints from weak lensing mass maps than the two-point functions. We present the cosmological results with a CNN from the KiDS-450 tomographic weak lensing dataset, constraining the total matter density $\Omega_m$, the fluctuation amplitude $\sigma_8$, and the intrinsic alignment amplitude $A_{\rm{IA}}$. We use a grid of N-body simulations to generate a training set of tomographic weak lensing maps. We test the robustness of the expected constraints to various effects, such as baryonic feedback, simulation accuracy, different value of $H_0$, or the lightcone projection technique. We train a set of ResNet-based CNNs with varying depths to analyze sets of tomographic KiDS mass maps divided into 20 flat regions, with applied Gaussian smoothing of $\sigma=2.34$ arcmin. The uncertainties on shear calibration and $n(z)$ error are marginalized in the likelihood pipeline. Following a blinding scheme, we derive constraints of $S_8 = \sigma_8 (\Omega_m/0.3)^{0.5} = 0.777^{+0.038}_{-0.036}$ with our CNN analysis, with $A_{\rm{IA}}=1.398^{+0.779}_{-0.724}$. We compare this result to the power spectrum analysis on the same maps and likelihood pipeline and find an improvement of about $30\%$ for the CNN. We discuss how our results offer excellent prospects for the use of deep learning in future cosmological data analysis.
Andrei Mesinger (Scuola Normale Superiore, Pisa)
Title:
The cosmic 21-cm signal: preparing for the Big Data revolution
Abstract:
The cosmic 21-cm signal is set to revolutionize our understanding of the first billion years of our Universe. Sensitive to the ionization and thermal state of the intergalactic medium, as well as to cosmological parameters, the 21-cm signal contains a wealth of physical insight. The upcoming Square Kilometre Array (SKA) in Australia will give us a 3D image of the epochs of cosmic dawn and reionization. The patterns of this image encode the properties of the unseen first generations of galaxies. The challenge is to interpret this enormous data set. Traditional approaches rely on the power spectrum statistic to infer galaxy properties from the large-scale images. However, the signal is highly non-Gaussian: using only the power spectrum wastes potentially valuable information. I will discuss how Convolutional Neural Networks (CNNs) have been used to infer galaxy parameters directly from 21-cm images: allowing the CNN to adaptively select its own summary statistic. Preliminary results of parameter recovery were promising, though only with comparable accuracy compared with MCMC approaches using the power spectrum. Most of the limitation we believe was due to our limited ability to tune the CNN hyper-parameters. I will discuss how we are performing a parameter exploration of hyper-parameters, in order to optimize the CNN's ability to extract parameters.
Matthew Ho (Carnegie Mellon University)
Title:
A Robust and Efficient Deep Learning Method for Dynamical Mass Measurements of Galaxy Clusters
Abstract:
Utilizing galaxy cluster abundance in precision cosmology requires large, well-defined cluster samples and robust mass measurement methods. In addition, modern cluster measurement techniques are expected to place a strong emphasis on efficiency and automation, as the wealth of detailed cluster data is expected to greatly increase with current and upcoming surveys such as DES, LSST, WFIRST, Euclid, and eROSITA. In this talk, I will discuss how we can leverage the use of deep learning models to infer dynamical cluster masses from spectroscopic samples with high precision and computational efficiency. I will demonstrate the ability of Convolutional Neural Networks (CNNs) to mitigate systematics in the virial scaling relation and produce dynamical mass estimates of galaxy clusters, using projected galaxies, with remarkably low bias and scatter. I will then discuss the performance of these methods relative to other leading analytic and machine learning dynamical mass estimators. Lastly, I will discuss our ongoing work in quantifying uncertainties in CNN mass predictions and our applications on spectroscopic datasets from the SDSS and GAMA surveys.
Roberto Trotta (Imperial College London)
Title:
A novel solution to supervised classification with biased training sets
Abstract:
Supervised classification in the presence of a biased training set is a widespread problem in cosmology, and will become even more important in the near future, when ML methods will become indispensable to deal with the large amount of data generated by observatories like the LSST. Training sets are often biased as a consequence of hard-to-avoid selection effects. Using such biased training sets blindly within Machine Learning methods leads to poor performance on the realistic test sets. I will present a novel solution, called STACCATO, that is able to produce classification results that are almost indistinguishable from what would be obtained using the "gold standard" of an unbiased training set. I will demonstrate the method on a realistic Supernova Type Ia classification challenge, and outline its potential applications to future problems in cosmology.
Wojciech Samek (Frauenhofer HHI Berlin)
Title:
Deep Understanding of Deep Models
Abstract:
Deep neural networks (DNNs) are reaching or even exceeding the human level on an increasing number of complex tasks. However, due to their complex non-linear structure, these models are usually applied in a black box manner, i.e., no information is provided about what exactly makes them arrive at their predictions. This lack of transparency is a major drawback in the sciences, where understanding is often more important than the prediction. This talk will demonstrate the effectivity of explanation techniques such as Layer-wise Relevance Propagation (LRP) when applied to various datatypes (images, text, audio, video, EEG/fMRI signals) and neural architectures (ConvNets, LSTMs), and will summarize what we have learned so far by peering inside deep models.
Christoph Schaefer (EPFL)
Title:
CNN based Strong Gravitational Lens Finder for Euclid
Abstract:
Future large-scale surveys with high resolution imaging will provide us with a few $10^5$ new strong galaxy-scale lenses. These strong lensing systems however will be contained in large data amounts which are beyond the capacity of human experts to visually classify in a unbiased way. The talk will present how convolutional neural networks can be used to classify lenses, how well they perform and the challenges they still face.
Slides
Dezső Ribli (Eötvös Loránd University)
Title:
Galaxy shape measurement with convolutional neural networks
Abstract:
Weak gravitational lensing by large scale structure is a powerful probe of the dark sector of the universe in the era of large sky surveys. Cosmic shear is measured via galaxy shape estimation, which is notoriously hard for the faint and small galaxies which dominate the observations. Parametric fitting of galaxy surface brightness profiles was successfully applied in multiple recent WL surveys, however, the approach has some shortcomings. Simplified model profiles cannot fully capture the complexity of real galaxy morphologies, and the fitting procedure often does not converge for small and faint galaxies. Model fitting also has a significant computational burden which is expected to increase by orders of magnitudes for Euclid or LSST. Recently, convolutional neural networks revolutionized the field of computer vision, and applications in astrophysics are emerging too.
We train a convolutional neural network (CNN) to predict the shapes of galaxies from Dark Energy Survey (DES) DR1 imaging data using “ground truth” shape measurements from an overlapping superior, deeper survey with less sky coverage, the Canada-France Hawaii Telescope Lensing Survey (CFHTLenS). We demonstrate that CNN predictions from DES images reproduce the results of CFHTLenS at bright magnitudes and show a significantly higher correlation with CFHTLenS at fainter magnitudes than DES Y1 results with forward model fitting. Prediction of shape parameters with a CNN is also extremely fast, it takes only 0.2 milliseconds per galaxy, improving 4 orders of magnitude over model fitting. Our proposed setup is applicable to any survey with dedicated deep fields.
Tuesday 11th June
Barnabás Póczos (Carnegie Mellon University)
Title:
Bayesian Active Learning for Posterior Estimation and Experiment Design
Abstract:
In the first part of the talk we study active posterior estimation in a Bayesian setting when the likelihood is expensive to evaluate. Existing techniques for posterior estimation are based on generating samples representative of the posterior. Such methods do not consider efficiency in terms of likelihood evaluations. In order to be query efficient we treat posterior estimation in an active regression framework. We propose myopic query strategies to choose where to evaluate the likelihood and implement them using Gaussian processes. Via experiments on a series of synthetic and real examples we demonstrate that this approach is significantly more query efficient than existing techniques and other heuristics for posterior estimation. In the second part we propose Bayesian bandit methods for black-box optimization. These methods are used in a variety of applications including hyper-parameter tuning and experiment design. Recently, multi-fidelity methods have garnered considerable attention since function evaluations have become increasingly expensive in such applications. Multi-fidelity methods use cheap approximations to the function of interest to speed up the overall optimization process. However, most multi-fidelity methods assume only a finite number of approximations. In many practical applications however, a continuous spectrum of approximations might be available. For instance, when tuning an expensive neural network, one might choose to approximate the cross validation performance using less data N and/or few training iterations T. Here, the approximations are best viewed as arising out of a continuous two dimensional space (N,T). In this work, we develop a Bayesian optimization method, BOCA, for this setting. We demonstrate the applicability of these methods in cosmology and other scientific problems.
Slides
Elena Giusarma (Flatiron Institute CCA)
Title:
Predicting neutrino simulations with convolutional neural network
Abstract:
Cosmology plays a crucial role in the investigation of neutrino properties since it is sensitive to the absolute scale of neutrino masses. Understanding the role of those particles in our Universe is one of the most important challenges in modern cosmology. Upcoming large scale structure surveys are expected to have enough statistical power to be able to detect the mass of neutrinos. In light of a large amount of new cosmological available in the next decade, the need for non-standard neutrino simulations will be essential for making theoretical predictions, for generating mock data and for analyzing a large amount of data. Producing such simulations requires the usage of significant computational resources for an extended period of time; thus the development of new computational methods are needed to accelerate this process. In this talk, I will propose a new approach to predict fast non-standard cosmological simulations with massive neutrinos by applying deep learning models and I will analyze the effects of those particles on standard simulations.
Tilman Troester (University of Edinburgh)
Title:
Painting with baryons: augmenting N-body simulations with gas using deep generative models
Abstract:
Running hydrodynamical simulations to produce mock data of large-scale structure and baryonic probes, such as the thermal Sunyaev-Zeldovich (tSZ) effect, at cosmological scales is computationally challenging. We propose to leverage the expressive power of deep generative models to find an effective description of the large-scale gas distribution and temperature. We train two deep generative models, a variational auto-encoder and a generative adversarial network, on pairs of matter density and pressure slices from the BAHAMAS hydrodynamical simulation. The trained models are able to successfully map matter density to the corresponding gas pressure. We then apply the trained models on 100 lines-of-sight from SLICS, a suite of N-body simulations optimised for weak lensing covariance estimation, to generate maps of the tSZ effect. The generated tSZ maps are found to be statistically consistent with those from BAHAMAS. We conclude by considering a specific observable, the angular cross-power spectrum between the weak lensing convergence and the tSZ effect and its variance, where we find excellent agreement between the predictions from BAHAMAS and SLICS, thus enabling the use of SLICS for tSZ covariance estimation. ArXiv:1903.12173
Fernando Perez-Cruz (Swiss Data Science Center)
Title:
Improved Bi-GAN with marginal likelihood equalization
Abstract:
We propose a novel training procedure for improving the performance of generative adversarial networks (GANs), especially to bidirectional GANs. First, we enforce that the empirical distribution of the inverse inference network matches the prior distribution, which favors the generator network reproducibility of the seen samples. Then, we propose a scalable reformulation of the method by Wu et al. to estimate the marginal log-likelihood (LL) of any sample. We have found that the marginal LL per sample over training and validation sets shows a severe overrepresentation of a certain type of samples. Finally, to address this issue, we propose to retrain the bidirectional GAN using non-uniform sampling for the mini-batch selection, resulting in improved FID, smaller reconstruction error, and higher average LL.
Davide Piras (University College London)
Title:
A novel approach in generating high-resolution cosmological maps using machine learning
Abstract:
Being able to produce thousands of simulations of our Universe with increasing precision is a challenging task, and the need for faster generative methods is today compelling. Generative machine learning algorithms have shown surprisingly promising results in training on cosmological images, and then generate new data points which can be described by the same statistical distribution as the input data. On the other hand, cosmological data allow for a validation of machine learning algorithms since, unlike natural images as the ones that are part of the most widely used datasets, it is possible to test the statistical information carried by the images beyond simple visual verification. In this work, we present a method to train a Variational Auto-Encoder using data from high-resolution Gaussian and lognormal random field maps on the sphere. In order to deal with the non-Euclidean domain, we suitably transform our data using the spherical harmonic expansion and rearranging the harmonic coefficients into square images. We perform a suite of statistical tests on the generated maps, and we are able to validate our approach with high confidence. Moreover, we explore how the latent variables are used when we mix maps with different cosmological parameters in our dataset. Finally, we describe how this method could be extended to obtain a parametric emulator of full 3-D simulations of our Universe.
Ben Moews (Institute for Astronomy, University of Edinburgh)
Title:
Synthetic datasets for modern cosmology: Creating galaxies with multi-stage GANs
Abstract:
Astronomy of the 21st century increasingly finds itself with extreme quantities of data. This growth is ripe for modern technologies such as deep learning, as deep image processing techniques have the potential to allow astronomers to automatically identify, classify, segment and deblend various astronomical objects. Since galaxies are a prime contender for such applications, we explore the staged use of generative adversarial networks (GANs), a class of generative models, to produce physically realistic galaxy images. In cosmology, such datasets can aid in the calibration of shape measurements for weak lensing by augmenting data with synthetic images, e.g. in upcoming surveys like LSST and Euclid. By measuring the distributions of multiple physical properties, we show that images generated with our approach closely follow the distributions of real galaxies, further establishing state-of-the-art GAN architectures as a valuable tool for modern-day astronomy.
Nathanaël Perraudin (Swiss Data Science Center)
Title:
Using generative adversarial networks to produce n-body simulations: a discussion of the challenges and current solutions
Abstract:
N-body simulations play a fundamental role in the field of cosmology, but they often require heavy computations. Generative adversarial networks (GANs) have been recently proposed as a solution to lower the computational complexity for generating such simulations. Although several promising results have recently appeared in the literature, at least two fundamental practical issues remain. 1) How to handle the data dimensionality when extending from 2D to 3D data? 2) How to learn a distribution conditioned on cosmological parameters? In this talk, we discuss these two challenges, propose some solutions and show some initial results…
Posters:
- Farida Farsian "Foreground models recognition through Neural Networks" poster
- Tomasz Kacprzak "Fast Point Spread Function Modeling with Deep Learning" poster
- Tommaso Ronconi "To build a mockingbird - Realistic a posteriori halo populator" poster
- Hector Javier Hortua Orjuela "Estimation of Cosmological Parameters via ConvNets" poster
Michaël Defferrard (EPFL)
Title:
DeepSphere: Efficient spherical Convolutional Neural Network with HEALPix sampling for cosmological applications
Abstract:
Convolutional Neural Networks (CNNs) are a cornerstone of the Deep Learning toolbox and have led to many breakthroughs in Artificial Intelligence. So far, these neural networks (NNs) have mostly been developed for regular Euclidean domains such as those supporting images, audio, or video. Because of their success, CNN-based methods are becoming increasingly popular in Cosmology. Cosmological data often comes as spherical maps, which make the use of the traditional CNNs more complicated. The commonly used pixelization scheme for spherical maps is the Hierarchical Equal Area isoLatitude Pixelisation (HEALPix). We present a spherical CNN for analysis of full and partial HEALPix maps, which we call DeepSphere. The spherical CNN is constructed by representing the sphere as a graph. Graphs are versatile data structures that can represent pairwise relationships between objects or act as a discrete representation of a continuous manifold. Using the graph-based representation, we define many of the standard CNN operations, such as convolution and pooling. With filters restricted to being radial, our convolutions are equivariant to rotation on the sphere, and DeepSphere can be made invariant or equivariant to rotation. This way, DeepSphere is a special case of a graph CNN, tailored to the HEALPix sampling of the sphere. This approach is computationally more efficient than using spherical harmonics to perform convolutions. We demonstrate the method on a classification problem of weak lensing mass maps from two cosmological models and compare its performance with that of three baseline classifiers, two based on the power spectrum and pixel density histogram, and a classical 2D CNN. Our experimental results show that the performance of DeepSphere is always superior or equal to the baselines. For high noise levels and for data covering only a smaller fraction of the sphere, DeepSphere achieves typically 10\% better classification accuracy than the baselines. Finally, we show how learned filters can be visualized to introspect the NN. Code and examples are available at https://github.com/SwissDataScienceCenter/DeepSphere.
Nicoletta Krachmalnicoff (SISSA)
Title:
Estimation of cosmologial parameters from large scale CMB polarized maps with CNN:
Abstract:
In this talk I will describe a novel method for the application of Convolutional Neural Networks (CNNs) to fields defined on the sphere, using the HEALPix tessellation scheme. In a recent paper (arXiv:1902.04083) we have developed a pixel-based approach to implement convolutional layers on the spherical surface, similarly to what is commonly done for CNNs in Euclidian space. The main advantage of this algorithm is its simplicity and the fact that it is fully integrable with existing, highly optimized, libraries for NNs (e.g., PyTorch or TensorFlow). Our approach could therefore be of great interest for the Astrophysics and Cosmology communities. I will present the application of such a code to CMB data analysis, and, in particular, preliminary results on cosmological parameter estimation directly from CMB maps. Our results show that CNNs can reach an accuracy comparable to standard spectrum-based bayesian methods when estimating parameters affecting the very large scales of CMB temperature maps, for full or partial sky coverage. We have tested the CNN also on polarized maps, and proved, for the first time, that the network is able to distinguish between polarization E and B-modes. An important application of such a method is related to the estimation of the optical depth at reionization from CMB maps (tau parameter). In my talk I will also show preliminary results on this point, demonstrating that CNNs have the potential to perform similarly to the standard, commonly used, methods and, therefore, could become a valid tool to crosscheck results.
Motritz Munchmeyer (Perimeter Institute)
Title:
Fast Wiener filtering with neural networks and applications to CMB lensing
Abstract:
We show how a neural network can be trained to perform extremely fast approximate Wiener filtering of CMB maps. We propose an innovative neural network architecture, which guarantees linearity in the data map. Our method does not require Wiener filtered training data, but rather learns Wiener filtering from tailored loss functions which are mathematically guaranteed to be minimized by the exact solution. Wiener filtering is the computational bottleneck in many optimal CMB analyses, including lensing, and we discuss how neural networks can be used to improve them. The paper (work together with Kendrick Smith) will appear in the next weeks.
Denitsa Staicova (Institute of Nuclear Research and Nuclear Energy, Bulgarian Academy of Sciences)
Title: The multi-measure cosmological model and its peculiar effective potential
Abstract:
The multi-measures model applied to cosmology has been recently shown to reproduce qualitatively the expected stages of the Universe evolution, along with some unexpected features. In this article, we continue its exploration with a detailed study of the effective potential of the model. An analysis of the limits of applicability of the effective potential show that during most of the Universe evolution, the effective potential is a very good approximation of the actual potential term. There is, however, a deviation between the two occurring in the earliest moments of the evolution, which has an important role on the behavior of the inflaton scalar field. In the studied cases, during this initial time, the inflaton is increasing in absolute value, which seems consistent with "climbing up" the effective potential. To investigate this behavior, we use numerical integration to find the numerical potential and we show that during the early stage of the evolution, the shape of the numerical potential is very different from that of the effective one and instead of a left plateau followed by steep slope, one observes only a slope with additional local maximum and minimum. This result demonstrates that for complicated equations of motion of the inflaton, one should not rely only on the notion of kinetic and effective potential terms to describe the problem as they may not be accurate in the entire numerical domain.
Slides
Wednesday 12th June
Tom Charnock (Institut d'Astrophysique de Paris)
Title:
Lossless non-linear data compression for cosmological surveys
Abstract:
Using novel machine learning techniques and statistical training, I will show how we can extract the most informative summaries of cosmological data given only simulations of this data. In essence, we build a neural network which compresses the data by training it to maximise the Fisher information. Since the method depends on calculating the Fisher information, simulations are only needed at a single fiducial parameter value making the necessary number of simulations extremely small compared to other machine learning methods. The summaries from this information maximising neural network are unbiased-estimators of the model parameters, allowing us to get first order approximate constraints on these model parameters. We can also perform approximate Bayesian computation to approach the true posterior distribution of parameters. The information maximising neural network is essential for future cosmological data analysis: it can be used to directly probe the informational content of features in the cosmological surveys, or perhaps more interestingly, to find a single additional summary which allows us to extract more constraining power from the surveys than current summaries, such as the power spectrum, etc.
Markus Michael Rau (Carnegie Mellon University)
Title:
Bayesian Optimization for scalable Bayesian photometric redshift inference
Abstract:
In the era of large area photometric surveys like DES, KiDS and LSST it becomes increasingly important to marginalize jointly over cosmological parameters and systematics like photometric redshift uncertainty. I present a Bayesian inference framework to efficiently infer photometric redshift distributions for large samples of galaxies from their photometric and clustering information. I will present our recent work on a hierarchical photometric redshift pipeline for cross-correlation inference based on logistic Gaussian processes (https://arxiv.org/abs/1904.09988). Going further I will showcase how photometric redshift inference based on SED templates and spectroscopic calibration data can be included into a joint redshift posterior. Finally, I will discuss how this posterior redshift uncertainty can be efficiently and accurately marginalized over in a cosmological analysis.
Luca Tortorelli (ETH Zürich, IPA)
Title:
The Luminosity Function measurement with Approximate Bayesian Computation
Abstract:
The galaxy Luminosity Function (LF) is a key observable not only for galaxy formation and evolution studies, but it has a dominant role for cosmology too, since it holds fundamental informations about the power spectrum of the primordial density fluctuations. Here, we propose an innovative technique to forward model wide-field broad and narrow-band galaxy surveys, using our fast image simulator (UFig) and consequently measure the LF of galaxies. Relying on the method described in Herbel+ 2017, we use Approximate Bayesian Computation to adjust the galaxy population model parameters of the simulations to match the observations from a range of imaging and spectroscopic surveys such as CFHTLS, Subaru-HSC and PAUS (Tortorelli+ 2018, JCAP 11, 035). By minimizing the distance between the datasets based on a Random Forest Classifier, we can obtain constraints on the LF of blue and red galaxies as a function of redshift, as well as other galaxy physical properties, without the limitations coming from the determination of individual galaxies photometric redshifts.
Keir Rogers (Oskar Klein Centre for Cosmoparticle Physics, Stockholm University)
Title:
Optimised interpolation of cosmological simulations: Bayesian emulator optimisation
Abstract:
In order to test cosmological theories, we must compare data to accurate models of our observations (i.e., we must evaluate the likelihood function). However, this is often impossible because the necessary cosmological simulations are too computationally expensive for the millions of likelihood samples we need (e.g., using MCMC methods). I will present a supervised learning solution which combines Gaussian process emulation with Bayesian optimisation of the training set. Here, a small training set of simulations is used to predict simulation outputs throughout parameter space (thus allowing evaluation of the likelihood) using a Gaussian process model. This is made accurate and efficient using Bayesian optimisation to build up the training set, conditional on the information already learnt and the uncertainty present from previous iterations of the training data. I will demonstrate an application of cosmological inference from the Lyman-alpha forest (a spectroscopic probe of the intergalactic medium at redshifts 2 to 6) at BOSS precision using hydrodynamical simulations. It is found that this new technique improves the precision of power spectrum estimates which propagates to improved constraints on cosmological parameters (an order of magnitude reduction in the 95% credible region) with 15% fewer simulations than by not using Bayesian optimisation. This method will also be important for the many other emulators of the cosmic large-scale structure — e.g., galaxy clustering, weak lensing or 21 cm.
Benjamin Giblin (University Of Edinburgh)
Title:
Accurately probing gravity on non-linear scales: machine learning our way beyond LCDM
Abstract:
Cosmic emulators, employing machine learning to predict observables given a training set of numerical simulations, have proven invaluable in calibrating models on small-scales and improving cosmological constraints. However, to date, emulators have been trained to predict only a small number of modest extensions to vanilla LCDM cosmology, which strongly limits our capacity to test the standard model.
We present a solution to this problem, having developed an emulator which, combined with the “halo model response” methodology from Cataneo et al. (2018), facilitates computationally inexpensive non-linear matter power spectra predictions in arbitrary cosmological models, including LCDM, with accuracies at the per cent level. We find that one requires only ~100 standard dark-matter-only LCDM simulations, with modified initial conditions, on which to train our emulator. Hence, we demonstrate that machine learning facilitates a significant step forward in practically constraining dark energy and modified gravity paradigms.
Davide Gualdi (Institute of Cosmos Sciences, University of Barcelona)
Title:
Artificial intelligence applied to the modelling of the large scale structure statistics
Abstract:
The first part of the talk will focus on using A.I. for the modelling of the matter power spectrum when a few cosmological parameters are let free to vary. The goal is to drastically speed up the running-time of MCMC sampling of posterior distributions requiring the recomputation of the matter power spectrum for the evaluation of the likelihood function (which is usually done by using either CAMB or CLASS). This at the same time also reduces the computational resources needed to run such pipelines, passing from CPU’s clusters to a single processor. The second part of the talk will focus on the modelling of the bispectrum using A.I., exploring its potential to accurately model the statistics at non-linear scales, where perturbation theory struggles or makes it unfeasable to compute the data-vector in a reasonable amount of time (to run MCMC-sampling).
Deborah Bard (NERSC, Lawrence Berkeley National Lab)
Title:
AI at the exascale: challenges and opportunities for cosmology
Abstract:
It is no coincidence that the rise of AI as a valuable tool for science has come at an interesting time for computing, where the end of Moore’s Law has meant that energy constraints are increasingly driving hardware innovation. AI is playing a growing role in shaping computing architecture, even at the largest scales. The US Department of Energy (DOE) recently announced the USA's first exascale supercomputers, coming in 2021/22 at a combined cost of $1.1B. These both contain GPU architectures, and have been designed specifically for large-scale AI science applications that will be partly developed by the US DOE ExaLearn program, a new co-design center for exascale machine learning technologies. In addition, many other specialized energy-efficient architectures designed for AI workloads have emerged in the past few years, including FPGAs, custom ASICs like Google’s TPU and Graphcore, and neuromorphic machines. In this talk I will explore the interplay between energy-efficient hardware and the rise of AI as a serious factor in scientific computing. I will focus on the challenges and opportunities for AI in the exascale era, with a focus on applications to cosmology.