Lossless, Scalable Implicit Likelihood Inference for Cosmological Fields

We present a comparison of simulation-based inference to full, field-based analytical inference in cosmological data analysis. To do so, we explore parameter inference for two cases where the information content is calculable analytically: Gaussian random fields whose covariance depends on parameters through the power spectrum; and correlated lognormal fields with cosmological power spectra. We compare two inference techniques: i) explicit field-level inference using the known likelihood and ii) implicit likelihood inference with maximally informative summary statistics compressed via Information Maximising Neural Networks (IMNNs). We find that a) summaries obtained from convolutional neural network compression do not lose information and therefore saturate the known field information content, both for the Gaussian covariance and the lognormal cases, b) simulation-based inference using these maximally informative nonlinear summaries recovers nearly losslessly the exact posteriors of field-level inference, bypassing the need to evaluate expensive likelihoods or invert covariance matrices, and c) even for this simple example, implicit, simulation-based likelihood incurs a much smaller computational cost than inference with an explicit likelihood. This work uses a new IMNNs implementation in $texttt{Jax}$ that can take advantage of fully-differentiable simulation and inference pipeline. We also demonstrate that a single retraining of the IMNN summaries effectively achieves the theoretically maximal information, enhancing the robustness to the choice of fiducial model where the IMNN is trained.

Catalog-free modeling of galaxy types in deep images: Massive dimensional reduction with neural networks

Current models of galaxy evolution are constrained by the analysis of catalogs containing the flux and size of galaxies extracted from multiband deep fields carrying inevitable observational and extraction-related biases which can be highly correlated. In practice, taking all of these effects simultaneously into account is difficult, and derived models are inevitably biased. To address this issue, we use robust likelihood-free methods for the inference of luminosity function parameters, made possible via massive compression of multiband images using artificial neural networks. This technique makes the use of catalogs unnecessary when comparing observed and simulated multiband deep fields and constraining model parameters. A forward modeling approach generates galaxies of multiple types depending on luminosity function parameters and paints them on photometric multiband deep fields including both the instrumental and observational characteristics. The simulated and the observed images present the same selection effects and can therefore be properly compared. We train a fully-convolutional neural network to extract the most model-parameter-sensitive summary statistics out of these realistic simulations, shrinking down the dimensionality of the summary space. Finally, using the trained network to compress both observed and simulated deep fields, the model parameter values are constrained through Population Monte Carlo likelihood-free inference. Using synthetic photometric multiband deep fields similar to the CFHTLS and D1/D2 deep fields and massively compressing them through the convolutional neural network, we demonstrate the robustness, accuracy and consistency of this new catalog-free inference method. We are able to constrain the parameters of luminosity functions of different types of galaxies and our results are fully compatible with the classic catalog extraction approaches.

Bayesian Neural Networks

In recent times, neural networks have become a powerful tool for the analysis of complex and abstract data models. However, their introduction intrinsically increases our uncertainty about which features of the analysis are model-related and which are due to the neural network. This means that predictions by neural networks have biases which cannot be trivially distinguished from being due to the true nature of the creation and observation of data or not. In order to attempt to address such issues we discuss Bayesian neural networks: neural networks where the uncertainty due to the network can be characterised. In particular, we present the Bayesian statistical framework which allows us to categorise uncertainty in terms of the ingrained randomness of observing certain data and the uncertainty from our lack of knowledge about how data can be created and observed. In presenting such techniques we show how errors in prediction by neural networks can be obtained in principle, and provide the two favoured methods for characterising these errors. We will also describe how both of these methods have substantial pitfalls when put into practice, highlighting the need for other statistical techniques to truly be able to do inference when using neural networks.

Detecting outliers in astronomical images with deep generative networks

With the advent of future big-data surveys, automated tools for unsupervised discovery are becoming ever more necessary. In this work, we explore the ability of deep generative networks for detecting outliers in astronomical imaging datasets. The main advantage of such generative models is that they are able to learn complex representations directly from the pixel space. Therefore, these methods enable us to look for subtle morphological deviations which are typically missed by more traditional moment-based approaches. We use a generative model to learn a representation of expected data defined by the training set and then look for deviations from the learned representation by looking for the best reconstruction of a given object. In this first proof-of-concept work, we apply our method to two different test cases. We first show that from a set of simulated galaxies, we are able to detect $sim90%$ of merging galaxies if we train our network only with a sample of isolated ones. We then explore how the presented approach can be used to compare observations and hydrodynamic simulations by identifying observed galaxies not well represented in the models.

Super-resolution emulator of cosmological simulations using deep physical models

We present an extension of our recently developed Wasserstein optimized model to emulate accurate high-resolution features from computationally cheaper low-resolution cosmological simulations. Our deep physical modelling technique relies on restricted neural networks to perform a mapping of the distribution of the low-resolution cosmic density field to the space of the high-resolution small-scale structures. We constrain our network using a single triplet of high-resolution initial conditions and the corresponding low- and high-resolution evolved dark matter simulations from the Quijote suite of simulations. We exploit the information content of the high-resolution initial conditions as a well constructed prior distribution from which the network emulates the small-scale structures. Once fitted, our physical model yields emulated high-resolution simulations at low computational cost, while also providing some insights about how the large-scale modes affect the small-scale structure in real space.

Neural physical engines for inferring the halo mass distribution function

An ambitious goal in cosmology is to forward-model the observed distribution of galaxies in the nearby Universe today from the initial conditions of large-scale structures. For practical reasons, the spatial resolution at which this can be done is necessarily limited. Consequently, one needs a mapping between the density of dark matter averaged over ~Mpc scales, and the distribution of dark matter halos (used as a proxy for galaxies) in the same region. Here we demonstrate a method for determining the halo mass distribution function by learning the tracer bias between density fields and halo catalogues using a neural bias model. The method is based on the Bayesian analysis of simple, physically motivated, neural network-like architectures, which we denote as neural physical engines, and neural density estimation. As a result, we are able to sample the initial phases of the dark matter density field whilst inferring the parameters describing the halo mass distribution function, providing a fully Bayesian interpretation of both the initial dark matter density distribution and the neural bias model. We successfully run an upgraded BORG inference using our new likelihood and neural bias model with halo catalogues derived from full N-body simulations. We notice orders of magnitude improvement in modelling compared to classical biasing techniques.

Painting halos from cosmic density fields of dark matter with physically motivated neural networks

We present a novel halo painting network that learns to map approximate 3D dark matter fields to realistic halo distributions. This map is provided via a physically motivated network with which we can learn the non-trivial local relation between dark matter density field and halo distributions without relying on a physical model. Unlike other generative or regressive models, a well motivated prior and simple physical principles allow us to train the mapping network quickly and with relatively little data. In learning to paint halo distributions from computationally cheap, analytical and non-linear density fields, we bypass the need for full particle mesh simulations and halo finding algorithms. Furthermore, by design, our halo painting network needs only local patches of dark matter density to predict the halos, and as such, it can predict the 3D halo distribution for any arbitrary simulation box size. Our neural network can be trained using small simulations and used to predict large halo distributions, as long as the resolutions are equivalent. We evaluate our models ability to generate 3D halo count distributions which reproduce, to a high degree, summary statistics such as the power spectrum and bispectrum, of the input or reference realizations.

Fast likelihood-free cosmology with neural density estimators and active learning

Likelihood-free inference provides a framework for performing rigorous Bayesian inference using only forward simulations, properly accounting for all physical and observational effects that can be successfully included in the simulations. The key challenge for likelihood-free applications in cosmology, where simulation is typically expensive, is developing methods that can achieve high-fidelity posterior inference with as few simulations as possible. Density-estimation likelihood-free inference (DELFI) methods turn inference into a density estimation task on a set of simulated data-parameter pairs, and give orders of magnitude improvements over traditional Approximate Bayesian Computation approaches to likelihood-free inference. In this paper we use neural density estimators (NDEs) to learn the likelihood function from a set of simulated datasets, with active learning to adaptively acquire simulations in the most relevant regions of parameter space on-the-fly. We demonstrate the approach on a number of cosmological case studies, showing that for typical problems high-fidelity posterior inference can be achieved with just $mathcal{O}(10^3)$ simulations or fewer. In addition to enabling efficient simulation-based inference, for simple problems where the form of the likelihood is known, DELFI offers a fast alternative to MCMC sampling, giving orders of magnitude speed-up in some cases. Finally, we introduce textsc{pydelfi} -- a flexible public implementation of DELFI with NDEs and active learning -- available at url{https://github.com/justinalsing/pydelfi}.