⚛️ phenomenology

NAE, Statistically

This paper introduces the Normalized Autoencoder (NAE) to provide a statistically interpretable, probabilistic framework for neural anomaly detection in new physics searches, demonstrating its validity through toy models, jet analysis, and Bayesian uncertainty quantification.

Original authors: Ranit Das, Jonathan Ostertag-Henning, Tilman Plehn, Lorenz Vogel

Published 2026-08-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ranit Das, Jonathan Ostertag-Henning, Tilman Plehn, Lorenz Vogel

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

At the Large Hadron Collider, a massive machine buried beneath the border of France and Switzerland, physicists smash protons together at nearly the speed of light. The goal is to find signs of new, unknown physics that might explain the universe's deepest mysteries. For decades, the strategy has been to guess what a new particle might look like, build a computer model of that guess, and then hunt for it in the collision data. But what if the new physics looks nothing like our guesses? What if it is hidden in a way we haven't imagined? In recent years, scientists have turned to artificial intelligence to act as a guide. Instead of hunting for a specific target, these computer programs learn what "normal" looks like based on the vast majority of data, and then flag anything that looks strange or unusual. This approach is powerful, but it has a major flaw: the computer gives a score saying something is weird, but it cannot tell the physicist how much to trust that score. Without a way to measure the reliability of the result, the search remains scientifically incomplete.

A team of researchers at the University of Heidelberg has developed a new method to fix this problem. They created a system that not only spots the unusual events but also calculates the statistical certainty of its own judgment. The core of their work involves a type of neural network called an autoencoder. Imagine a machine that tries to compress a complex image into a small summary and then reconstruct the image from that summary. If the machine is trained only on normal data, it becomes very good at rebuilding normal things. When it sees something strange, it fails to reconstruct it properly, and the error in the reconstruction becomes a signal that something is new. However, the researchers realized that simply measuring this error was not enough. They needed to connect that error to a probability, a mathematical way of saying how likely it is that an event belongs to the normal crowd.

To solve this, the team introduced a "normalized" version of the autoencoder. This system treats the reconstruction error not just as a mistake, but as a measure of energy, where low energy means the event is normal and high energy means it is rare. By carefully training the network to understand the total landscape of possibilities, they ensured that the energy score directly corresponds to a probability. This means the computer can now say, "This event is 99% likely to be normal," or "This event is highly unusual." To make this even more robust, they added a layer of uncertainty to the system. By allowing the network's internal settings to vary slightly during training, the system learns not just a single answer, but a range of possible answers. This gives the scientists a built-in error bar, showing them how confident the computer is in its own assessment.

The researchers first tested this idea on a simple, two-dimensional model where they already knew the correct answer. They found that the system successfully learned the true distribution of the data, accurately identifying the most common areas and the rare edges. When they added the uncertainty feature, the system correctly identified that it was less sure about the data points far away from the center, where it had seen fewer examples. The computer's confidence matched the reality of the data: it was very sure in the crowded center and appropriately cautious in the empty corners. They also tested a method where two of these systems were trained on different types of data and compared against each other. This dual setup allowed them to calculate the ratio of probabilities between two different scenarios, effectively mimicking a more complex, supervised search without needing to know the answer in advance.

Encouraged by these results, the team applied their method to real-world data from the Large Hadron Collider. They used a dataset of "jets," which are sprays of particles created when quarks or gluons are knocked loose during a collision. Specifically, they looked for top quark jets hidden among a sea of ordinary quantum chromodynamics jets. In this high-stakes environment, the system performed impressively. It successfully distinguished the rare top quark jets from the common background, and the uncertainty estimates provided a clear picture of the system's reliability. The researchers found that the computer's learned energy differences closely matched the results of a traditional, supervised classifier, which is the gold standard for these searches. This confirmed that the new method was not just guessing; it was learning the underlying physics of the data.

The study also revealed that the system behaves differently depending on what it is looking for. When trained to find top quarks, the system was very consistent. When trained to find ordinary jets hidden among top quarks, the uncertainty was higher, reflecting the greater difficulty of the task. This is a crucial insight, as it shows the system is honest about its limitations. The researchers demonstrated that by adjusting a specific parameter in the training process, they could align the system's energy scores with the known probabilities from a supervised classifier. This alignment suggests that the method can be tuned to work effectively even in the complex, high-dimensional space of real particle collisions.

Ultimately, this work bridges a critical gap between the raw power of artificial intelligence and the rigorous demands of statistical physics. It transforms anomaly detection from a black box that produces a mysterious score into a transparent tool that provides a probability and a measure of confidence. By proving that these networks can learn the true likelihood of events and quantify their own uncertainty, the researchers have provided a new, reliable way to search for the unknown. The method does not require a specific theory of what new physics might look like; it simply learns what is normal and flags the rest, with a clear statement of how sure it is. This capability could fundamentally change how physicists analyze the flood of data from the Large Hadron Collider, allowing them to spot the unexpected with a level of statistical rigor that was previously out of reach.

Technical Summary: Statistically Interpretable Anomaly Detection with Normalized Autoencoders

Problem Statement
The search for physics beyond the Standard Model (BSM) at the Large Hadron Collider (LHC) increasingly relies on unsupervised anomaly detection to identify rare events without committing to specific signal hypotheses. While modern machine learning techniques, particularly bottleneck-based autoencoders (AEs), offer a pathway to flag anomalous events, they suffer from a critical limitation: the lack of a rigorous statistical interpretation. The standard reconstruction error (Mean Squared Error, MSE) used as an anomaly score does not correspond to a learned likelihood, making it difficult to assign statistical significance or quantify uncertainties to the detection of new physics. This paper addresses two specific questions: (1) What is the statistical interpretation of the anomaly score in a normalized autoencoder (NAE)? and (2) Can this interpretation be extracted with a reliable uncertainty estimate?

Methodology
The authors propose a framework grounded in energy-based models to bridge the gap between autoencoder reconstruction errors and probabilistic likelihoods.

Normalized Autoencoder (NAE): The authors define the reconstruction error $|x - x'|^2$ as an energy function $E_\theta(x)$ . By relating this energy to a probability density via the Boltzmann distribution $p_\theta(x) \propto e^{-E_\theta(x)/T}$ , the NAE is trained to minimize the negative log-likelihood rather than just the MSE. To handle the intractable partition function $Z_\theta$ required for likelihood normalization, the authors employ Langevin Monte Carlo (LMC) sampling. They utilize "on-manifold initialization," where samples are generated in the latent space and decoded to feature space, allowing for efficient exploration of the low-energy regions of the model distribution.
Bayesian NAE (BNAE): To address statistical uncertainties, particularly in sparse data regimes, the authors extend the NAE to a Bayesian framework. By replacing deterministic network weights with probability distributions (specifically Gaussian variational distributions), the BNAE learns a distribution over the energy function. This allows for the calculation of the mean and variance of the learned log-likelihood, providing a direct measure of the uncertainty in the anomaly score.
Dual-NAE Test: In realistic scenarios where the true underlying density is unknown, the authors introduce a "dual-NAE" test. Two NAEs are trained on different datasets (e.g., two different signal regions or background types). The difference in their learned, temperature-rescaled energies is compared against the log-likelihood ratio derived from a supervised classifier trained to distinguish the two datasets. This serves as a proxy to validate whether the unsupervised NAEs have correctly learned the underlying densities.

Key Contributions and Results

Toy Model Validation: Using a two-dimensional toy model with a known true density (a mixture of two Gaussians), the authors demonstrate that the NAE successfully learns the target density, achieving a Pearson correlation of 0.979 between the learned and true log-likelihoods. The Bayesian extension (BNAE) further provides meaningful uncertainty estimates, showing that the learned uncertainties are best calibrated in the region around the two Gaussian peaks and are underestimated as one moves towards the corners of the distribution.
Dual-NAE Verification: In the toy model, the log-likelihood ratio constructed from two separate NAEs closely matches the ratio from a supervised classifier (Pearson correlation of 0.987), confirming that the NAEs learn the correct relative likelihoods even without explicit supervision.
Jet Tagging Application: The methodology is applied to the TOPTAGGING dataset (top jets vs. QCD jets).
- Performance: The NAE resolves the asymmetry often seen in standard AEs, where AEs perform poorly when reconstructing complex signals from simple backgrounds (or vice versa). The NAE achieves an Area Under the Curve (AUC) of 0.9036 for top-tagging and 0.8071 for QCD-tagging, outperforming standard AEs.
- Uncertainty Quantification: The BNAE provides uncertainty bands on the ROC curves. While the BNAE performance is compatible with the NAE baseline, it exhibits larger uncertainties for the more challenging QCD-tagging task, reflecting the difficulty of learning the density in high-dimensional, sparse jet image spaces.
- Statistical Interpretation: Applying the dual-(B)NAE test to jet images, the authors find that the learned energy differences approximately reproduce the log-likelihood ratio of a supervised classifier (Pearson correlations of 0.851 for NAE and 0.878 for BNAE). However, unlike the toy model, achieving a slope of $m=1$ requires careful tuning of the temperature parameter $T$ , suggesting that learning the correct temperature in high-dimensional spaces remains a challenge.

Significance and Claims
The paper claims that the Normalized Autoencoder provides a necessary probabilistic interpretation for autoencoder-based anomaly detection, tying the anomaly score directly to a learned likelihood. By introducing the Bayesian NAE, the authors demonstrate that it is possible to extract this likelihood with an associated uncertainty estimate, moving anomaly detection from a heuristic scoring system toward a statistically rigorous tool.

The authors emphasize that while the NAE successfully learns the density in low-dimensional toy models and shows promise in high-dimensional jet tagging, the practical application in LHC analyses requires careful handling of the temperature parameter and the calibration of uncertainties, particularly in the tails of distributions where data is sparse. The work establishes a foundation for statistically interpretable, unsupervised searches for new physics, though it acknowledges that fully automated extraction of the correct temperature for high-dimensional feature spaces remains an open problem.

Technical Summary: Statistically Interpretable Anomaly Detection with Normalized Autoencoders

More like this