💻 computer science

The Reconstruction Gap: Boundary Noise and the Limits of Holographic Bulk Recovery in Deep Learning, with Implications for Black-Box AI Auditing

This paper introduces the "reconstruction gap" to demonstrate that holographic deep learning models suffer from significant and unpredictable internal metric errors under noisy boundary conditions, thereby quantifying the fundamental limitations of black-box AI auditing in certifying a model's internal mechanisms.

Original authors: Alfredo Merlet

Published 2026-10-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Alfredo Merlet

Original paper licensed under CC BY 4.0 (https://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 a universe where everything happening inside a vast, three-dimensional room is perfectly encoded on the two-dimensional walls that surround it. This is the core idea of a famous concept in theoretical physics known as holography. Just as a hologram on a credit card stores a three-dimensional image in a flat surface, this theory suggests that the complex interior of space and time, including gravity and black holes, is written entirely on its boundary. For decades, physicists have used this idea to solve difficult equations, but a newer, more experimental approach has emerged: using artificial intelligence to read the interior from the boundary. In this setup, a computer network acts as a translator, trying to reconstruct the hidden geometry of the universe's interior based solely on data from the edge.

This method has become a popular way to test how well machines can learn the laws of physics. However, a critical question has been largely ignored: what happens when the data on the boundary is not perfect? In the real world, measurements are always slightly noisy, and sensors are never flawless. A new study by researcher Alfredo Merlet investigates exactly this problem. He set up a controlled digital laboratory where the "true" interior of a simulated universe was known with absolute certainty. He then trained artificial intelligence networks to guess that interior using boundary data that was sometimes clean and sometimes corrupted by random noise. The goal was to see how much of the true interior the machines could actually recover, and whether their perfect performance on the outside guaranteed they understood the inside.

The results reveal a startling disconnect between how a system behaves and what it actually knows. When the AI was trained on perfect, noise-free data, it performed almost flawlessly at the boundary, correctly classifying nearly 99 percent of the test cases. To an outside observer, the machine appeared to have mastered the physics. Yet, when the researchers looked inside the machine's learned model, they found that the reconstruction of the deep interior was far from perfect. Even in the best-case scenario, the error in the region closest to the center of the simulated universe was nearly 30 percent. The machine had learned to mimic the surface behavior perfectly while failing to capture the true structure of the deep interior.

The situation deteriorates rapidly when noise is introduced. The study found that a tiny amount of noise, equivalent to just one percent error in the boundary measurements, caused the error in the deep interior to nearly double. This happened while the machine's overall accuracy on the surface dropped by only a tiny fraction. It is a case where the machine looks excellent from the outside but has fundamentally misunderstood the core mechanics. The researchers tried to fix this by increasing the amount of data the machine studied, but simply adding more data did not help; the error in the deep interior remained stubbornly high. They also tried using a standard mathematical technique to smooth out the machine's learning, which helped near the boundary but actually made the deep interior worse once the noise level passed a certain threshold.

Perhaps the most significant finding concerns the limits of auditing these systems. In the world of artificial intelligence, regulators and auditors often have only "black-box" access, meaning they can see what goes in and what comes out, but they cannot see the internal weights and calculations. The study simulated this by creating hundreds of different AI networks that all performed identically well on the boundary data. To an auditor looking only at the output, these machines were indistinguishable. However, their internal structures were wildly different. Some of these "identical" machines had internal errors that varied by a factor of more than three. Even more concerning, when these machines were tested on new, unseen situations, their performance diverged by as much as 42 percentage points. One machine might work perfectly on a new task while another, which looked exactly the same during testing, would fail completely.

The study concludes that looking at the output of a system is not enough to know how it works. The researchers found that even when they could see the internal structure of the machines, a single global measure of error could not predict which machine would fail on new data. This suggests that for high-stakes applications, simply checking if a model works on standard tests is insufficient. To truly understand a system, one must ask specific, targeted questions about the parts of the mechanism that will be used in the real world. The study does not claim that all AI models are holographic universes, but it uses this unique physical setup to prove a hard truth: a system can be perfectly right at the edge and completely wrong at the center, and without looking inside, there is no way to tell the difference.

Technical Summary: The Reconstruction Gap in Holographic Deep Learning

Problem Statement
The paper addresses a critical gap in the literature regarding "holographic deep learning," a field applying the AdS/CFT correspondence (Maldacena's conjecture) to machine learning. In this framework, a neural network's layers represent radial slices of an anti-de Sitter (AdS) spacetime, and its weights represent the bulk metric. While existing literature (e.g., Hashimoto et al., 2018a; Filev, 2026) has established that networks can recover bulk geometry from boundary data, it has largely assumed ideal conditions. The specific problem identified is the lack of understanding regarding how much interior information is lost when boundary data is observed imperfectly (i.e., with noise) and whether different internal structures can produce identical boundary behaviors. This is framed as an inverse problem with implications for AI auditing, where auditors often have only "black-box" access (inputs/outputs) and must infer internal mechanisms ("the bulk").

Methodology
The study employs a controlled computational laboratory using a known ground truth to measure the "reconstruction gap"—the relative error between the learned metric and the true metric, resolved by radial region.

Model Architecture: The design follows Hashimoto et al. (2018a), discretizing a 4D planar AdS–Schwarzschild black hole. A scalar field with mass $m^2 = -1$ and quartic self-interaction propagates from the boundary ( $\eta \to \infty$ ) to the horizon ( $\eta = 0$ ). The network consists of $N$ layers, where each layer's single trainable weight corresponds to the metric function $h(\eta)$ at that radial slice.
Learning Task: The network is trained to classify boundary conditions as "regular" (field remains finite at the horizon) or "irregular" based on boundary inputs $(\phi, \pi)$ . The true metric is $h(\eta) = 3 \coth(3\eta)$ .
Experimental Design:
- Experiment 1 (Factorial Grid): 900 training runs varying four factors: boundary noise levels (0% to 20% of input standard deviation), sample sizes (500 to 10,000), network depths (10, 20, 40 layers), and the presence of smoothness regularisation.
- Experiment 2 (Observational Equivalence): 400 networks trained from different random initialisations on a fixed, clean dataset to identify networks with indistinguishable in-distribution accuracy but potentially different internal structures.
Metrics:
- Reconstruction Gap ( $G_n$ ): Relative error $| \hat{h}_n - h(\eta_n) | / h(\eta_n)$ averaged over boundary, middle, and near-horizon regions.
- Behavioral Accuracy: Classification accuracy on clean test sets.
- Out-of-Distribution (OOD) Accuracy: Performance on boundary data drawn from regions not seen during training.

Key Results

Radial Structuring of Error: Even with clean data, the reconstruction gap is non-uniform. While boundary regions are recovered with high fidelity (~~1.7% error), the near-horizon region retains a significant error (~~30%) despite the network achieving 99.1% classification accuracy.
Asymptotic Sensitivity to Noise: Boundary noise disproportionately affects the deep interior.
- A 1% noise level reduces accuracy by only 0.5 percentage points but nearly doubles the near-horizon reconstruction error (from ~30% to ~55%).
- At 5% noise, accuracy remains high (~94.6%), but the near-horizon error exceeds 100% (the learned metric sign-reverses).
Ineffectiveness of Standard Remedies:
- Sample Size: Increasing data by 20-fold (500 to 10,000 points) does not reduce the near-horizon error, suggesting the issue is bias (systematic error from noise) rather than variance.
- Depth: Deeper networks do not improve structural fidelity in clean conditions; they simply extend the region of high error closer to the horizon.
- Regularisation: Smoothness regularisation reduces boundary error in clean conditions but worsens near-horizon reconstruction when noise exceeds ~6–7%.
Observational Equivalence and Underspecification:
- Networks with identical in-distribution accuracy (within 1%) can have internal metrics differing by a factor of three.
- These structurally distinct networks exhibit vastly different OOD behaviors, with accuracy gaps of up to 42 percentage points.
- Crucially, a global measure of structural error (the reconstruction gap) does not predict which network will fail on OOD data.

Significance and Claims
The paper positions its findings as a formal analogy rather than an equivalence, arguing that holographic deep learning serves as a unique laboratory for quantifying the limits of black-box auditing.

For Physics: The results argue that holographic reconstructions using real or experimental data must report region-specific uncertainty, particularly near the horizon, as standard benchmarks on clean data mask deep structural failures.
For AI Auditing: The study provides quantitative evidence for three propositions:
1. Behavioral Underdetermination: High in-distribution accuracy provides no information about the proximity of the internal mechanism to the truth (underspecification).
2. Deployment Risk: Observationally equivalent models can diverge significantly on out-of-distribution inputs, a risk that black-box testing alone cannot detect.
3. Limits of White-Box Access: Merely inspecting weights (white-box access) reveals structural dispersion but does not automatically predict failure modes. Effective auditing requires asking targeted questions about specific internal mechanisms relevant to the deployment context.

The author explicitly states that these are hypotheses derived from a controlled setting with known ground truth, intended to motivate further research into interpretable architectures with known internal targets, rather than direct claims about current deployed AI systems.

Technical Summary: The Reconstruction Gap in Holographic Deep Learning

More like this