🤖 machine learning

A Probabilistic Circuit-Induced Pseudo-Metric for Out-of-Distribution Detection

This paper introduces a novel out-of-distribution detection framework that leverages the internal hierarchical structure of Probabilistic Circuits to define a Hierarchical Likelihood Vector and a corresponding pseudo-metric, enabling unsupervised detection and shift localization without requiring held-out in-distribution data.

Original authors: Bhumika K, Vidhya S, Narayanan C Krishnan

Published 2026-08-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bhumika K, Vidhya S, Narayanan C Krishnan

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 security guard at a very exclusive club. Your job is to decide who gets in and who doesn't. Usually, you have a mental list of what your regulars look like: their height, their favorite hat, the way they walk. If someone walks in wearing a neon clown suit, you know immediately they don't belong. But what if the person looks almost like a regular, just with a slightly different shirt? Or what if the club is so huge that you can't remember every single detail of every regular? This is the challenge of "Out-of-Distribution" (OOD) detection in the world of Artificial Intelligence. It's the task of teaching a computer to say, "Hey, this data looks weird; it doesn't match what I learned," without needing a second list of examples to compare it against.

To understand how this paper tackles that problem, we need to look at the tool they are using: a "Probabilistic Circuit." Think of this not as a messy neural network, but as a giant, organized family tree of probabilities. At the bottom, you have simple facts (like "is this pixel black?"). As you move up the tree, these facts are combined into bigger ideas (like "is this a circle?"). The top of the tree gives you a final score: "How likely is this picture to be a cat?" The problem with the old way of doing things is that security guards (or AI models) often only look at that final score at the very top. They ignore all the interesting details in the middle of the tree. This paper argues that by looking at the whole tree, not just the top, we can catch the "clown suits" much better.

The researchers, working at the Indian Institute of Technology Palakkad, introduce a new method called the Hierarchical Likelihood Vector (HLV). Instead of just asking the AI, "Is this a cat?" (which gives a single number), they ask the AI a whole bunch of questions at different levels of its family tree. "Is this a circle?" "Is this a black pixel?" "Is this a specific type of curve?" They collect all these answers into a list, or a vector. This list is the HLV. It's like giving the security guard a checklist of every detail, from the shoes to the hat, rather than just a single "looks like a regular" stamp.

Using this checklist, they created a new way to measure distance called the Hierarchical Likelihood Distance (HLD). Imagine you have a "perfect regular" checklist and a "new guy" checklist. The HLD measures how far apart these two lists are. If the new guy's list is very different from the perfect regular's list, the distance is big, and the AI knows to sound the alarm. The clever part of this paper is that they figured out how to calculate the "perfect regular" checklist mathematically, directly from the AI's brain, without needing to save a bunch of extra photos of regulars to compare against later. This makes the system much more efficient and ready for real-world use where you might not have extra data lying around.

When they tested this idea on tabular data (like spreadsheets of customer information) and images of handwritten numbers (MNIST), the results were promising. In their simulations, the HLV method was better at spotting the "clowns" than the old methods that only looked at the top of the tree. For example, on the tabular datasets, when they tested batches of 100 items, the new method detected the weird data about 93% of the time, while the old methods were often stuck around 50% or lower. Even more importantly, the new method was more stable; it didn't get confused as easily when the data got slightly messy or when the AI's model wasn't perfect.

The paper also showed that this method can act like a detective. Because the HLV keeps track of every node in the tree, the researchers could look at which specific questions in the checklist were different. If the AI was confused about a digit '6' looking like a '0', the method could point out exactly which part of the image (the middle loop) was causing the trouble. This "localization" helps us understand why the AI thinks something is weird, rather than just knowing that it is.

However, the authors are careful to note that this isn't a magic bullet that solves everything forever. Their method relies on the AI model being a specific type of "Probabilistic Circuit" that is smooth and structured. If the model is messy or unstructured, the math doesn't work as cleanly. Also, while the method is great at spotting data that doesn't fit the model, if the model itself is a bad representation of reality (like a blurry photo of a cat), the system might still flag real cats as weird. The authors found that on high-resolution images (28x28 pixels), the false alarm rate went up because the model struggled to capture every tiny detail perfectly.

In short, this paper suggests that by listening to the whole conversation inside an AI's brain, rather than just the final conclusion, we can build better security systems for data. It turns a single "yes/no" decision into a rich, multi-layered investigation, allowing the AI to catch subtle shifts in data that other methods miss, all without needing a backup list of examples to compare against. It's a step toward making AI more aware of its own limits and more capable of spotting the unexpected.

Technical Summary: A Probabilistic Circuit-Induced Pseudo-Metric for Out-of-Distribution Detection

Problem Statement

Out-of-distribution (OOD) detection aims to determine whether data encountered at deployment is consistent with the distribution represented by a learned model. In the unsupervised setting, generative models are a natural choice as they explicitly model the in-distribution (ID). However, existing approaches face significant limitations:

The Likelihood Paradox: Scalar likelihoods from deep generative models (e.g., VAEs, Normalizing Flows) can assign higher scores to OOD samples than ID data.
Information Loss: Current Probabilistic Circuit (PC) based methods often reduce the entire circuit hierarchy to a single scalar likelihood at the root, discarding the richer probabilistic information encoded in internal nodes.
Data Dependency: Many distribution comparison methods (e.g., Maximum Mean Discrepancy) or calibration techniques require held-out ID reference samples or calibration sets, which are unavailable in strict deployment scenarios where the trained model is the sole representation of the ID.

The central question addressed is: Can the hierarchical probabilistic information encoded within a trained PC be used to construct a statistically calibrated, distribution-level OOD detector without relying on held-out ID data?

Methodology

The proposed framework introduces two core concepts: the Hierarchical Likelihood Vector (HLV) and the Hierarchical Likelihood Distance (HLD).

1. Hierarchical Likelihood Vector (HLV)

Instead of evaluating an observation $x$ solely at the root of the PC, the method evaluates it at a selected set of nodes $\mathcal{N} = \{n_1, \dots, n_m\}$ .

Definition: For a smooth and decomposable PC, the HLV is defined as $L_C(x) = (\ell_{n_1}(x), \dots, \ell_{n_m}(x))$ , where $\ell_n(x)$ is the likelihood of the observation restricted to the scope of node $n$ .
Significance: This vector preserves probabilistic summaries across multiple variable scopes. Input nodes provide variable-level summaries, while sum and product nodes provide summaries over progressively larger scopes. If the root is included, the conventional likelihood is retained alongside these hierarchical summaries.

2. Hierarchical Likelihood Distance (HLD)

The HLD measures the discrepancy between the learned distribution $P$ and a test distribution $Q$ by comparing their expected HLVs.

Definition: $d_C(P, Q) = \|\mu_P - \mu_Q\|_2$ , where $\mu_P = \mathbb{E}_{X \sim P}[L(X)]$ and $\mu_Q = \mathbb{E}_{X \sim Q}[L(X)]$ .
Theoretical Property: The authors prove that HLD is an Integral Probability Metric (IPM) over a function class $\mathcal{F}_C$ induced directly by the PC. Specifically, $\mathcal{F}_C$ consists of linear combinations of the node likelihood functions. Unlike generic IPMs where the function class is learned implicitly, every function in $\mathcal{F}_C$ has a direct probabilistic interpretation derived from the circuit structure.

3. Goodness-of-Fit Test and Analytic Threshold

The method formulates OOD detection as a goodness-of-fit hypothesis test ( $H_0: P = Q$ vs. $H_1: P \neq Q$ ).

Test Statistic: Given a test batch $D_Q$ of size $T$ , the statistic is $\hat{d}_C(P, Q) = \|\mu_P - \hat{\mu}_Q\|_2$ , where $\hat{\mu}_Q$ is the empirical mean of the HLVs in the batch.
Exact Computation: A key innovation is that for structured-decomposable PCs, the population mean $\mu_P$ and the covariance matrix $\Sigma_P$ of the HLV can be computed exactly and tractably from the trained circuit alone. This is achieved using the "Scope Trichotomy" property of structured-decomposable PCs, which ensures that the scopes of any two nodes are either disjoint, identical, or nested. This allows for exact marginalization and product operations without needing ID reference samples.
Decision Threshold: Under the null hypothesis, the test statistic follows a scaled generalized chi-square distribution. The authors derive an approximate closed-form threshold $\tau_\alpha$ by matching the first two moments (mean and variance) of this distribution, computed directly from $\text{tr}(\Sigma_P)$ and $\text{tr}(\Sigma_P^2)$ . This eliminates the need for held-out ID data for calibration.

Key Contributions

HLV and HLD: Introduction of a representation that captures hierarchical probabilistic summaries and a corresponding pseudo-metric (HLD) that compares distributions via their expected HLVs.
IPM Characterization: Proof that HLD is a PC-induced Integral Probability Metric, grounding the detector in established statistical theory.
Tractable Exact Computation: Demonstration that for structured-decomposable PCs, the necessary population statistics (mean and covariance of the HLV) can be computed exactly from the model parameters. This enables an analytic decision threshold without held-out ID data.
Empirical Validation: Evaluation on tabular datasets (Adult, Covertype, Sensorless, Census-KDD, Connect-4) and binarized MNIST, showing improved performance over root-likelihood, uncertainty-based (TDI), typicality-based, and kernel-based (MMD) baselines.

Experimental Results

The framework was evaluated using a class-wise OOD protocol where one class serves as ID and others as OOD.

Detection Power: HLD consistently achieves higher detection power than baselines, particularly at small to moderate batch sizes ( $T=10$ to $T=100$ ). For example, on tabular datasets at $T=100$ , HLD reached a detection rate of 0.987, compared to 0.542 for Root Likelihood and 0.608 for Typicality.
Robustness to Model Misspecification: When tested on real held-out ID data (where the model may not perfectly fit the true distribution), HLD exhibited a significantly slower increase in False Positive Rate (FPR) as batch size grew compared to other methods. At $T=1000$ on tabular data, HLD's FPR was 0.178, whereas MMD reached 0.807 and Root Likelihood 0.446.
Localization: The node-wise decomposition of the test statistic allows for the localization of distribution shifts. By ranking nodes based on their contribution to the discrepancy, the method can identify specific variable scopes (e.g., specific image regions in MNIST) responsible for the shift.
Ablation Studies: Combining leaf and sum nodes for the HLV provided the most robust performance across datasets. Adding product nodes did not consistently improve detection power and sometimes degraded it.

Significance and Claims

The paper claims that exploiting the hierarchical structure of Probabilistic Circuits offers a principled advantage over scalar-score methods. By treating the circuit not just as a likelihood calculator but as a generator of a rich, interpretable feature space (the HLV), the method constructs a distribution-level detector that is:

Self-Calibrating: It requires no held-out ID data for thresholding, relying instead on the exact structural properties of the trained PC.
Interpretable: It can localize the source of distribution shifts to specific nodes (variable scopes) within the circuit.
Robust: It maintains better calibration and detection power under model misspecification compared to existing baselines.

The authors acknowledge limitations: the statistical calibration is relative to the distribution represented by the trained PC (meaning poor model fit leads to higher FPR on real data), and the closed-form threshold relies on a Gaussian approximation which may be less accurate if the covariance spectrum is dominated by a few eigenvalues. Future work includes extending the framework to continuous random variables.