ltu-ili: ILI + Astro Papers

A repository for interesting papers on implicit likelihood inference (a.k.a. simulation-based inference or likelihood-free inference) and its applications in astronomy, astrophysics, and cosmology.

Machine Learning

Techniques

• 2023 - One-Line-of-Code Data Mollification Improves Optimization of Likelihood-based Generative Models

Neural Posterior/Likelihood/Ratio Estimation

• 2023 - Nicolas Chartier's tests of the sbi package, including:
  • Sensitivity Prior/Proposal Distribution
  • Architectures (MAF vs MDN)
  • Data vector length
• 2020 -

Bayesian Neural Networks

SBI Posterior Coverage / Calibration

• 2021 - Benchmarking Simulation Based Inference
• 2022 - Simulation-Based Calibration Checking for Bayesian Computation: The Choice of Test Quantities Shapes Sensitivity
• 2022 - Conditionally Calibrated Predictive Distributions by Probability-Probability Map: Application to Galaxy Redshift Estimation and Probabilistic Forecasting
• 2019 - Diagnostics for conditional density models and Bayesian inference algorithms

Interpretability

• 2022 - Efficient identification of informative features in simulation-based inference

Applications

Review Papers

Cosmology

Gravitational-Wave Astronomy

Neural Posterior Estimation

• 2023 - Neural Posterior Estimation with guaranteed exact coverage: the ringdown of GW150914
  • Using SNPE with Masked Autoregressive Flows
• 2023 - Flow Matching for Scalable Simulation-Based Inference
  • The paper builds upon the DINGO study and introduces Flow MAtching Posterior Estimation (FMPE)
  • "We represent x in frequency domain; for two LIGO detectors and complex f ∈ [20, 512] Hz, ∆f = 0.125 Hz, we have x ∈ R^{15744}": The authors do mention the use of an embedding network to compress x to 128 summaries but do not talk about SVD. They say the computation is roughly the same as DINGO.
• 2021 - Real-Time Gravitational Wave Science with Neural Posterior Estimation
  • M.Dax presents "DINGO" in the recording of his NeurIPS talk.
  • The paper presents an application of Group-equivarient Neural Posterior Estimation GNPE.
  • The pre-processing involves saving SVD representation of frequency-domain waveform on disk, and generating noise at training time with extrinsic parameters. Then, an embedding network precedes a normalizing flow. fmin = 20 Hz ,fmax = 1024 Hz , and ∆f = 0.125 Hz
  • Waveform model is IMRPhenomPv2. The training set comprises 5 million waveforms.
  • A particularity here, appart from the novelty of GNPE to treat "translation" parameters such as coalescence time, is that merger extrinsic parameters as well as noise realizations are randomly drawn during training.
• 2020 - Gravitational-wave parameter estimation with autoregressive neural network flows
  • Example of posterior inference on only 5 parameters at first.
  • Input data is 1 second of strain data with Fs = 1024Hz, whitened in frequency domain before IFFT and finally rescalling in the time domain (factor coming from the discretization of band-limited colored noise). * Waveform is IMRPhenomPv2.
  • 1 million data-parameter pairs, 90% for training and 10% for validation. The authors used a Conditional Variational AutoEncoder (CVAE) and a MAF.

Neural Ratio Estimation

• 2023 - Peregrine: Sequential simulation-based inference for gravitational wave signals
  • The authors apply (Sequential) Truncated Marginal Neural Ratio Estimation TMNRE.
  • Waveform model is SEOBNRv4PHM.
  • Both time and frequency-domain data are presented to the density estimator.
• 2020 - Lightning-Fast Gravitational Wave Parameter Inference through Neural Amortization
  • The input vectors of the network (Convolutional layers + Perceptron) are two real time-domain strains corresponding to two interferometers. Dimension is (2 × 8192) corresponding to H1/L1 strains.
  • The authors refer to this paper about the training data generation. Said paper deals with detection rather than inference.
  • Waveform model is IMRPhenomPv2

Galaxy Properties