⚛️ high-energy experiments

An Introduction to Bayesian and Frequentist Simulation-Based Inference with Machine Learning

This paper provides an overview of how machine learning-based simulation-based inference methods can be applied within both Bayesian and frequentist frameworks to solve inverse problems, while also addressing their use in Empirical Bayes and unfolding tasks, along with validation strategies and inherent limitations.

Original authors: Maximilian Dax, Theo Heimel, Gilles Louppe

Published 2026-07-27

📖 9 min read🧠 Deep dive

CC BY 4.0

Original authors: Maximilian Dax, Theo Heimel, Gilles Louppe

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but you can never see the culprit. All you have are the clues left behind at the scene: a muddy footprint, a broken window, or a strange smell. In the real world, scientists face this exact problem every day. They have a theory about how the universe works (the culprit), but they can only observe the data that theory produces (the clues). The challenge is to work backward from the clues to figure out the theory. This is called an "inverse problem."

Usually, scientists have a "forward map," a set of rules that tells them what clues to expect if their theory is true. But often, these rules are too messy or complex to write down as a simple math equation. Instead, they have to run a giant, complicated computer simulation to see what happens. This is where the trouble starts: if you can't write down the math, you can't easily use the standard detective tools to solve the case. You might have to guess, run the simulation a million times, and hope you get lucky. This is slow, expensive, and often impossible.

Enter a new kind of detective: Machine Learning. Instead of trying to solve the math puzzle every time a new clue appears, these AI detectives learn the pattern of the clues by studying millions of fake crime scenes generated by the computer. They learn to look at a clue and instantly guess what the culprit looks like. A new paper from the SciPost Physics Community Reports, titled "An Introduction to Bayesian and Frequentist Simulation-Based Inference with Machine Learning," acts as a friendly guidebook for these AI detectives. It explains how to teach them, how to trust them, and how to make sure they aren't just making things up.

The Two Ways to Think About Truth

Before we meet the AI, we need to understand the two different ways scientists usually think about "truth." The paper explains that there are two main schools of thought, and the AI needs to know which one you're using.

The first school is Bayesian. Imagine you are trying to guess the weight of a mystery box. You have a hunch based on past experience (maybe it feels like a cat, so it's probably around 10 pounds). This hunch is your "prior." Then, you get a new clue: the box makes a meowing sound. You update your guess. You combine your hunch with the new clue to get a "posterior" belief. In this view, the true weight of the box is a bit fuzzy; it's a range of possibilities that you get more confident about as you gather more clues. This is great when you have very few clues but a lot of background knowledge, like in astronomy where you might only see a star once.

The second school is Frequentist. Imagine you are a judge in a courtroom. You don't care about your hunches about the defendant. You only care about the evidence. If you ran this exact trial a million times with different random juries, would your verdict be right 95% of the time? In this view, the true weight of the box is a single, fixed number (even if we don't know what it is), and our job is to find a range of numbers that is guaranteed to catch the truth most of the time. This is the go-to method in particle physics, where they run the same experiment millions of times and want to be absolutely sure they aren't fooling themselves.

The paper shows that the new AI methods can play by the rules of both schools.

The AI Detective: Learning from Simulations

So, how does the AI detective work? The paper describes a process called Simulation-Based Inference (SBI).

Traditionally, if a scientist wanted to find the parameters of a model, they would need a formula that says, "If the parameters are X, the data will look like Y." But often, that formula is too hard to write down. They only have a simulator—a black box that takes parameters and spits out data.

The paper explains that instead of trying to reverse-engineer the black box, we can train a neural network (a type of AI) to be the reverse-engineer. Here is the trick:

The Training Phase: We run the simulator millions of times. We pick random parameters, run the simulation, and save the pair: (Parameters, Data). We do this over and over, creating a massive library of "fake" crime scenes and their culprits.
The Learning Phase: We feed this library to the AI. The AI learns the relationship between the data and the parameters. It learns to say, "When I see data that looks like this, the parameters were probably that."
The Inference Phase: Now, when a real scientist gets a new piece of data from the real world, they don't need to run the simulator again. They just feed the data to the trained AI, and poof—the AI gives them the answer instantly.

The paper highlights three main ways the AI learns this relationship:

Neural Posterior Estimation (NPE): The AI learns to draw a map of all possible answers. It doesn't just give one number; it draws a cloud of possibilities, showing you where the culprit is most likely to be hiding.
Neural Likelihood Estimation (NLE): The AI learns to predict how likely a specific set of parameters is to produce the data.
Neural Ratio Estimation (NRE): Instead of learning the whole map, the AI learns to compare two things. It answers the question: "Is this data more likely to come from suspect A or suspect B?" This is like a referee deciding which team is better without needing to know the exact score of every game.

The "Amortized" Advantage

One of the coolest things the paper points out is the concept of "amortized inference." Think of it like learning to ride a bike. The first time you try, it takes hours of falling and getting back up (the training phase). But once you've learned, you can ride anywhere instantly.

In traditional science, every time you get a new clue, you have to start from scratch, running thousands of simulations to figure out the answer. With this AI method, you pay the "learning cost" once, up front. After that, you can analyze a million different clues in the blink of an eye. This is a game-changer for fields like particle physics, where experiments generate millions of data points every second.

The "Gotchas": Validation and Limits

The paper is very careful not to say, "AI is magic, trust it blindly." In fact, a huge chunk of the paper is dedicated to validation—how do we know the AI isn't hallucinating?

Since the AI is trained on fake data, it might get really good at guessing the fake data but fail miserably when it sees something weird from the real world. The authors explain several ways to test the AI:

The "Two-Sample" Test: You give the AI a bunch of fake data and a bunch of real data (or data from a trusted method) and ask a simple classifier to tell them apart. If the classifier can easily tell them apart, the AI is bad. If the classifier is confused, the AI is doing a good job.
Calibration Checks: Imagine the AI says, "I'm 90% sure the answer is between 5 and 10." If you run this test a hundred times, the answer should actually be in that range 90 times. If it's only there 50 times, the AI is overconfident and lying.
Posterior Predictive Checks: You take the AI's best guess, run the simulator forward with that guess, and see if it produces data that looks like the real observation. If the AI's guess leads to a totally different kind of data, the guess is wrong.

The paper also warns about nuisance parameters. These are extra variables in the simulation that you don't care about but that mess up the results (like the temperature in a lab affecting a chemical reaction). The paper explains that the AI can handle these, but it makes the training much harder and more complex.

What the Paper Does Not Say

It is important to note what this paper does not claim. It does not say that AI has solved all scientific mysteries. It does not claim that these methods work perfectly for every single problem. The authors are very clear that these are tools that need to be validated carefully. They don't promise that the AI will work if the simulator itself is broken (if the "forward map" is wrong, the AI will just learn the wrong thing very quickly).

They also don't say that this replaces all other methods. Sometimes, if you have a simple math formula, the old-school methods are still fine. The AI is specifically for those "impossible" cases where the math is too hard to write down.

The Bottom Line

This paper is a roadmap for a new era of scientific discovery. It teaches us how to build AI detectives that can learn from complex simulations to solve inverse problems. It shows us how to train them using both Bayesian and Frequentist logic, how to combine many small clues into a big picture, and, most importantly, how to double-check their work so we don't get fooled.

The authors suggest that while the technology is powerful and fast, it is still young. We need better ways to validate these tools and better ways to handle the messy, real-world problems where the simulations might not be perfect. But for scientists drowning in data and stuck with black-box simulators, this guide offers a lifeline: a way to turn millions of computer simulations into instant, reliable answers.

Technical Summary: An Introduction to Bayesian and Frequentist Simulation-Based Inference with Machine Learning

Problem Statement

Inverse problems are ubiquitous in science and engineering, ranging from inferring model parameters from data to inverting detector effects. In many modern applications, the forward map from parameters $\mu$ to observations $x$ is known only implicitly via a simulator that can generate samples $x \sim p(x|\mu)$ , but the likelihood function $p(x|\mu)$ is intractable or computationally prohibitive to evaluate directly.

Traditional likelihood-based inference methods (e.g., Markov Chain Monte Carlo, nested sampling) fail in these "likelihood-free" scenarios. Conversely, classical simulation-based inference (SBI) methods, such as Approximate Bayesian Computation (ABC), often suffer from low acceptance rates or require hand-crafted summary statistics that lead to information loss. The challenge is to perform accurate parameter estimation (Bayesian or frequentist) and distribution-level inference (e.g., unfolding, empirical Bayes) efficiently, without requiring explicit likelihood evaluation, while maintaining rigorous statistical validity.

Methodology

The paper provides a comprehensive technical overview of using deep learning to solve these inverse problems, categorizing the approach into statistical frameworks and machine learning techniques.

1. Statistical Frameworks

The authors distinguish between two primary paradigms for handling uncertainty:

Bayesian Inference: Treats parameters $\mu$ as random variables with a prior $p(\mu)$ . The goal is to approximate the posterior $p(\mu|x)$ using Bayes' theorem. This framework is favored when prior knowledge exists or when dealing with single-event observations (common in astrophysics).
Frequentist Inference: Treats parameters as fixed and data as random. The goal is to construct confidence sets $C_{conf}(x)$ that guarantee coverage over repeated experiments. This is preferred in high-energy physics for discovery claims where prior-independent guarantees are required.
Distribution-Level Inference: Extends these concepts to inferring entire distributions $p(\mu)$ or $p(z)$ (latent space) rather than single parameters, applicable to tasks like detector unfolding and empirical Bayes.
Nuisance Parameters: The paper details how to handle parameters of no direct interest ( $\nu$ ) via marginalization (Bayesian) or profiling (Frequentist), noting the computational trade-offs in simulation-based settings.

2. Machine Learning Techniques for SBI

The core contribution is the application of deep generative models and classifiers to approximate the necessary distributions or ratios.

Density Estimation (Neural Posterior/Likelihood Estimation - NPE/NLE):
- NPE: Trains a conditional generative model $q_\theta(\mu|x)$ to approximate the posterior $p(\mu|x)$ . This is typically achieved by minimizing the forward Kullback-Leibler Divergence (KLD) between the true posterior and the model, using samples from the joint distribution $(\mu, x)$ . Normalizing flows are highlighted as a popular architecture due to their tractable density evaluation, though diffusion models and score-matching are also applicable.
- NLE: Trains a model $q_\theta(x|\mu)$ to approximate the likelihood. This requires an additional sampling step (e.g., MCMC) for Bayesian inference, making it less amortized than NPE.
- Training Dynamics: The paper emphasizes the importance of the "mass-covering" nature of the forward KLD, which ensures the model captures all modes of the posterior, contrasting it with the "mode-seeking" reverse KLD.
Ratio Estimation (Neural Ratio Estimation - NRE):
- Instead of learning densities directly, NRE learns the likelihood-to-evidence ratio $r(x|\mu) = p(x|\mu)/p(x)$ (or the posterior-to-prior ratio) using a binary classifier.
- The classifier is trained to distinguish between samples from the joint distribution $p(x, \mu)$ and samples from the product of marginals $p(x)p(\mu)$ .
- This approach offers architectural flexibility (no invertibility constraint) but requires combining the learned ratio with sampling algorithms (like MCMC) for full Bayesian inference.
Incorporating Likelihood Information:
- Importance Sampling (NPE-IS): If the likelihood is tractable, NPE outputs can be reweighted to yield asymptotically exact posterior samples and unbiased estimates of the Bayesian evidence.
- Latent Likelihoods: If a tractable likelihood exists for an intermediate latent step $z$ (e.g., $p(z|\mu)$ ), this information can be used to train ratio estimators via regression rather than classification, reducing gradient variance.
Data Modality and Amortization:
- Embedding Networks: For complex data (images, sets, time series), embedding networks $h_\psi(x)$ reduce data to summary statistics before conditioning the generative model.
- Combining Observations: For frequentist inference with many independent observations, likelihood ratios can be multiplied directly, allowing for fully amortized inference even with millions of data points. Bayesian combination requires sampling from the product of posteriors, which is less efficient.

3. Validation

The paper stresses that validation is critical but difficult, as the target distribution is often inaccessible. It outlines several necessary (but not sufficient) diagnostic tools:

Training Convergence: Monitoring loss and checking for overfitting (memorization of single pairs).
Comparison to Reference: Comparing against MCMC or nested sampling results where feasible.
Classifier Two-Sample Tests (C2ST): Using a classifier to detect differences between the learned distribution and a reference (or the joint truth) to identify failure modes.
Calibration Tests:
- PP Plots: Checking if the cumulative distribution of a sample under the learned model matches the uniform distribution.
- Coverage Plots: Verifying that the true parameter falls within the $\alpha$ -credible/confidence set with probability $\alpha$ .
- Classifier Calibration: Ensuring the output probabilities of the ratio estimator are well-calibrated.
Posterior Predictive Checks: Simulating data from the inferred posterior to see if the real observation is typical, which helps detect simulator misspecification.

Key Contributions

Unified Framework: The paper synthesizes Bayesian and frequentist perspectives within the context of machine learning-based SBI, demonstrating how the same neural architectures (NPE, NLE, NRE) can serve both paradigms.
Distribution-Level Extension: It explicitly connects parameter inference to distribution-level tasks (unfolding, empirical Bayes), showing how iterative or direct optimization methods can recover latent distributions.
Technical Deep Dive: It provides a detailed mathematical treatment of training objectives (forward vs. reverse KLD), the mechanics of amortization, and the specific challenges of handling nuisance parameters and combining observations.
Validation Standards: As part of the VERaiPHY initiative, the paper establishes a rigorous set of validation protocols (C2ST, calibration, predictive checks) tailored for SBI, emphasizing that no single test guarantees accuracy.

Results and Claims

The paper does not present new experimental results or a specific application case study. Instead, it serves as a technical introduction and review. Its "results" are the synthesis of existing methodologies and the demonstration of their theoretical properties:

It claims that amortized inference (training once, inferring many times) is a key advantage of deep SBI, enabling real-time analysis and handling of massive datasets.
It asserts that forward KLD training is generally more stable and robust for capturing multimodal posteriors than reverse KLD.
It highlights that importance sampling combined with NPE can provide asymptotically exact results and evidence estimation, bridging the gap between approximate and exact inference.
It notes that while classifier-based ratio estimation is flexible, it loses full amortization for Bayesian inference unless combined with sampling.

Significance and Outlook

The paper positions itself as a foundational resource for the VERaiPHY initiative, aiming to establish verification and validation standards for AI in physics. Its significance lies in:

Bridging Disciplines: It connects machine learning advances with statistical physics and high-energy physics requirements.
Addressing the Validation Gap: It argues that the lack of universal validation techniques is the primary barrier to the widespread adoption of SBI in scientific discovery.
Future Directions: It identifies open challenges, including the treatment of large numbers of nuisance parameters, robust estimation of Bayesian evidence, and handling simulator misspecification (where the simulator cannot generate the observed data).

The authors conclude that while SBI is becoming a standard tool, its adoption in producing new scientific results is still emerging. The field requires continued development of robust validation standards and methods to handle out-of-distribution data and misspecified simulators before it can be fully trusted for critical discovery claims.