PaperlayerView exact revision on arXiv

Source-linked AI summary

The frontier of simulation-based inference

Kyle Cranmer, Johann Brehmer, Gilles Louppe

arXiv:1911.01429v3stat.MLcs.LGstat.ME

TL;DR

Simulation-based inference addresses scientific inverse problems in which complex simulators provide high-fidelity models but intractable likelihoods. This review synthesizes traditional and emerging approaches, including machine-learning and probabilistic-programming developments, and describes a rapidly advancing frontier with potential to improve inference quality or replace heuristic methods with statistically grounded ones.

  • Problem

    Complex simulators define statistical models with likelihoods that are typically intractable because they integrate over large latent execution spaces, making inference difficult.

  • Method

    The paper reviews simulation-based and likelihood-free inference, emphasizing traditional methods, machine-learning developments, probabilistic programming, and workflows combining new directions.

  • Results

    The review identifies a rapidly advancing frontier that may yield significant improvements in inference quality or move scientific analyses from heuristics toward statistical methods tied to mechanistic models.

  • Takeaways & Limitations

    Simulation-based inference is expanding through machine learning, active learning, probabilistic programming, and differentiable programming, enabling work with high-dimensional data and new inference capabilities.

Abstract

from arXiv · show

Many domains of science have developed complex simulations to describe phenomena of interest. While these simulations provide high-fidelity models, they are poorly suited for inference and lead to challenging inverse problems. We review the rapidly developing field of simulation-based inference and identify the forces giving new momentum to the field. Finally, we describe how the frontier is expanding so that a broad audience can appreciate the profound change these developments may have on science.

1. Simulation-based inference

Simulation-based inference treats a simulator as the statistical model, generating latent states and observations from parameters. Because simulators vary widely, no single inference method fits every problem.

  • 1. Simulation-based inference: A simulator maps parameters θ through latent states z_i to observed data x, thereby defining the statistical model.Simulators may combine stochastic and deterministic steps, with continuous or discrete latent spaces and varying access to internal states.
  • 1. Simulation-based inference: There is no one-size-fits-all inference method because simulators differ in latent structure, dimensionality, differentiability, and observability.The review therefore emphasizes considerations for selecting an appropriate approach rather than prescribing one universal method.
  • 1. Simulation-based inference: Inference goals differ between estimating input parameters θ, latent variables z, or both, with nuisance parameters included when they affect observed-data distributions.The review focuses primarily on parametric inference for θ while also discussing methods for z.

B. Inference.

Simulation-based inference is difficult because the simulator-defined likelihood integrates over potentially enormous latent execution spaces, while Bayesian inference may add a high-dimensional evidence integral. The review examines methods that support frequentist or Bayesian inference despite these obstacles.

  • B. Inference.: Frequentist inference may target point estimates or confidence sets, whereas Bayesian inference typically targets a posterior for observed data and a given prior.In both paradigms, the likelihood p(x|θ) is a key ingredient.
  • B. Inference.: The simulator-defined likelihood is typically intractable because it integrates over all possible latent execution traces.For large latent spaces, this integral cannot be computed explicitly, obstructing both frequentist and Bayesian inference.
  • B. Inference.: Simulation-based or likelihood-free inference techniques enable frequentist or Bayesian inference despite intractable simulator likelihoods.The paper frames these methods as a specialization of inverse uncertainty quantification for accurate, stochastic simulators.
  • B. Inference.: Bayesian inference has a separate intractability source: the evidence integral over parameters θ becomes difficult in high-dimensional problems.Markov Chain Monte Carlo and variational inference are commonly used to address this additional challenge.
  • B. Inference.: The likelihood factorizes into individual terms for multiple i.i.d. observations, whereas time-series data are treated as a single high-dimensional observation.This distinction affects computational cost because i.i.d. inference may require repeated evaluations of individual likelihood terms.

C. Traditional methods.

Traditional simulation-based inference methods approximate inference through rejection based on simulated-data similarity or estimated likelihoods. Both approaches face dimensionality and efficiency limitations and therefore depend on informative low-dimensional summaries.

  • C. Traditional methods.: Rejection ABC samples θ from the prior, simulates x_sim, and retains θ when ρ(x_sim, x_obs) < ϵ.Accepted samples approximate the posterior, with ρ measuring simulated–observed-data distance and ϵ setting tolerance.
  • C. Traditional methods.: ABC becomes exact as ϵ → 0, but continuous-data acceptance vanishes and small tolerances require unfeasibly many simulations.Larger tolerances improve sample efficiency at the expense of inference quality.
  • C. Traditional methods.: ABC requires rerunning the inference algorithm for new observations because the data affect rejection and, in advanced methods, the proposal distribution.It is therefore best suited to a single observation or at most a few i.i.d. data points.
  • C. Traditional methods.: A second classical approach estimates the simulated-data likelihood with histograms or kernel density estimation, then performs frequentist or Bayesian inference as if it were tractable.This approach was widely used in particle-physics measurements.
  • C. Traditional methods.: Both traditional approaches suffer from the curse of dimensionality, with worst-case simulation requirements increasing exponentially in data dimension.They consequently rely on low-dimensional summary statistics whose information retention determines inference quality.

2. Frontiers of simulation-based inference

Traditional simulation-based inference faces shortcomings in sample efficiency, inference quality, and amortization. Recent progress centers on machine learning, active learning, and deeper integration of simulation code with inference workflows.

  • Traditional approaches suffer from poor sample efficiency, information loss from summary statistics, and limited amortization across new observations.ABC and classical density estimation can require prohibitively many simulations, discard information through low-dimensional summaries, and differ in whether expensive computation must be repeated.
  • New capabilities are intended to improve all three shortcomings of traditional simulation-based inference.
  • Machine learning enables inference with higher-dimensional data and neural-network surrogates that benefit from advances in deep learning.
  • Active learning improves sample efficiency by selecting simulator parameter points expected to increase knowledge.
  • Automatic differentiation, probabilistic programming, and simulator-derived information expose more of the simulator to the inference workflow.

A. A revolution in machine learning.

Machine learning expands simulation-based inference by supporting high-dimensional density estimation and flexible surrogate models, while active learning targets computational efficiency. These developments improve inference quality and sample efficiency but retain the simulator as a generative black box.

  • Neural density estimation models probability densities in high dimensions and can condition on inputs to represent likelihoods or posteriors.Normalizing flows provide tractable densities and generation, while conditional extensions model p(x|θ) or p(θ|x).
  • Autoregressive models factorize high-dimensional distributions into successive conditional densities and provide tractable density estimation and generation.
  • Generative adversarial networks use an adversary to train expressive generators whose densities are intractable.
  • Active learning iteratively uses knowledge from previous simulations to choose parameter points expected to increase knowledge.
  • Machine learning and active learning can improve inference quality and sample efficiency relative to classical methods without fundamentally changing the simulator–inference separation.

C. Integration and augmentation.

Integration and augmentation move simulation-based inference beyond treating simulators as black-box generators. Probabilistic programming, automatic differentiation, and simulator-derived quantities provide additional information for inference and surrogate training.

  • Deeper integration opens the simulator to inference and enables tighter coupling between simulation and inference.
  • Probabilistic programming adds random draws and conditioning on observations, enabling inference over simulator execution traces conditioned on observed data.Recent work can add these capabilities to existing simulators with minimal codebase changes.
  • Simulator-derived quantities can augment training data and dramatically increase sample efficiency for surrogate training.
  • Automatic differentiation efficiently calculates derivatives and provides a way to exploit deep-learning advances within existing simulation codes.
  • The integration of inference and simulation, together with augmented training data, has the potential to change how simulators are used in inference.

3. Workflows for simulation-based Inference

Simulation-based inference workflows combine simulator runs, optional data reduction, and inference or surrogate-model stages in different configurations. They differ in whether inference uses the simulator directly or a learned or estimated surrogate, and in what Bayesian output they provide.

  • Inference workflows assemble common building blocks into approaches that may resemble traditional ABC and density-estimation methods or differ fundamentally.
  • Simulator parameters are drawn from proposal distributions that may be static or iteratively chosen with active learning, and outputs may be used directly or reduced to summaries.
  • Direct methods compare simulator outputs with data, whereas surrogate methods train an estimation or machine-learning model on simulated data.
  • Algorithms address an intractable likelihood by constructing surrogates for the likelihood or likelihood ratio, or by replacing it implicitly with rejection probability.
  • Bayesian methods may provide posterior samples or a tractable function approximating the posterior, and may require choosing inferred quantities early or later.

A. Using the simulator directly during inference.

Direct simulator-based inference includes ABC, probabilistic programming, and related algorithms, with trade-offs involving summaries, tolerances, efficiency, and simulator structure. When the simulator is differentiable or its final stage is tractable, asymptotically exact Bayesian inference can avoid ABC’s main limitations.

  • ABC relies on low-dimensional summary statistics, while classifier ABC trains a classifier to estimate discrepancies between observed and simulated data.
  • Prior-based parameter proposals make rejection ABC sample-inefficient when much prior mass disagrees strongly with the observations.
  • Alternative algorithms improve ABC’s sample efficiency by simulating at parameter points expected to provide the greatest posterior information.
  • Differentiable simulators or tractable final simulator stages enable asymptotically exact Bayesian inference without distance tolerances or summary statistics.
  • Probabilistic programming requires a probabilistic-language simulator and either a tractable final likelihood or an ABC-like comparison.
  • These methods can infer both simulator input parameters and latent variables, providing physical interpretability and more efficient latent-process inference than standard ABC.
  • Direct simulator-based inference lacks amortization, so each new observation requires repeating the inference chain.

B. Surrogate models.

Surrogate models replace repeated direct simulator use with learned approximations of parameters, likelihoods, likelihood ratios, or posteriors. Their amortization enables efficient evaluation for new data, while the target quantity determines prior dependence and downstream inference requirements.

  • Surrogate-based inference amortizes computation: after upfront simulation and training, new data can be evaluated efficiently.This scaling is particularly favorable for many independent and identically distributed observations.
  • Direct parameter inversion produces point estimates, but their probabilistic interpretation in terms of likelihoods or posteriors is not obvious.
  • Conditional density estimators can learn either the posterior p(θ|x) or the likelihood p(x|θ) as a simulator surrogate.
  • Classifier-based approaches learn likelihood ratios such as p(x|θ0)/p(x|θ1) or p(x|θ0)/p(x) from simulated data.
  • All three surrogate classes require parameters of interest to be specified upfront and implicitly marginalize over other simulator variables.
  • Posterior learning directly targets Bayesian inference but is prior-dependent, whereas likelihood and likelihood-ratio learning support frequentist inference or model comparison and require MCMC or VI for Bayesian posterior samples.
  • Additional simulator-derived information can augment surrogate training data and improve sample efficiency in the inference task.
  • Surrogate architectures can incorporate inductive biases reflecting the data structure or the simulator’s causal structure.

C. Preprocessing and postprocessing.

Preprocessing and postprocessing extend core inference methods through learned summaries, inference compilation, calibration, and diagnostics. These choices depend on the data, simulator, inference goals, computational cost, and available simulator control, while model misspecification remains a limitation across prescribed and implicit models.

  • Learned summaries can replace hand-chosen low-dimensional summaries, which may lose information relative to the original data.
  • High-dimensional raw data are often compressed into moderate-dimensional high-level features before entering the inference workflow.
  • Inference compilation trains a neural network from initial simulator runs to perform sequential importance sampling over parameters and latent variables.
  • Parametric-bootstrap calibration can provide confidence sets and posteriors with proper coverage and credibility, but may require many simulations.
  • Post-inference diagnostics compare simulator and surrogate data, test known estimator expectations, vary invariant reference parameters, use ensembles, and assess asymptotic network-output properties.
  • Model misspecification affects both prescribed and implicit models and is commonly addressed by adding flexibility and nuisance parameters.
  • Choosing an approach depends on inference goals, dimensionalities, summary-statistic availability, simulator structure and cost, execution control, and accessible simulator quantities.

D. Recommendations.

The recommended method depends on which simulator quantities are available and whether the goal includes latent-variable inference. Where suitable summaries are absent, neural likelihood or likelihood-ratio surrogates are recommended, while active learning and amortization must be balanced against one another.

  • Differentiable simulators, accessible joint data–latent likelihoods, and probabilistic-programming models should use the corresponding powerful inference algorithms.
  • Probabilistic programming is recommended when inference must cover both parameters and latent variables.
  • Traditional techniques remain reasonable when powerful low-dimensional summaries are established; otherwise, neural likelihood or likelihood-ratio surrogates are recommended.
  • Likelihood-ratio learning is preferred over likelihood learning when generating synthetic data from the surrogate is unnecessary because it can leverage supervised learning methods.
  • Active learning can improve sample efficiency across inference techniques, but it trades data-specific efficiency against amortization across observations.The appropriate balance depends on the number of observations and posterior sharpness relative to the prior.

4. Discussion

Simulation-based inference is moving beyond expert-dependent traditional methods through machine learning, active learning, and tighter simulator integration. The review argues that this expanding frontier may improve inference quality or replace heuristics with statistically grounded approaches.

  • Discussion: Traditional ABC and classical density-estimation methods rely heavily on expert-designed summary statistics, limiting their usefulness for high-dimensional data.These approaches are labor intensive, and powerful summary statistics may not be obvious.
  • Discussion: Machine-learning methods enable inference directly from high-dimensional data and reduce reliance on manually designed summaries.
  • Discussion: Active learning connects statistical and computational considerations by using acquired knowledge to improve the sample efficiency of inference methods.
  • Discussion: The advancing frontier may bring significant inference-quality improvements or replace heuristics with statistical methods tied to the underlying mechanistic model.
Loading 1911.01429v3…