The problem with post hoc explanations is well documented in the machine learning literature. Methods such as LIME, SHAP, and Grad-CAM build simplified surrogate models that approximate a trained network’s behavior, but studies by Alvarez-Melis and Jaakkola, Laugel and colleagues, Slack and colleagues, and Cynthia Rudin have shown that these surrogates can diverge sharply from the underlying model. An explanation model may match the black box’s accuracy while relying on entirely different features, producing explanations that look plausible and are simply wrong. In high-stakes domains such as medical diagnosis and legal decision-making, the researchers argue, such misleading accounts are not merely inconvenient but dangerous. Rudin’s influential 2019 position paper called for abandoning the explain-the-black-box paradigm altogether in favor of models that are interpretable from the ground up, with every component designed for a distinct, human-understandable purpose.

NPCs answer that call by building on concept bottleneck models, an architecture introduced by Pang Wei Koh and colleagues in 2020. Concept bottleneck models split prediction into two stages: a neural network first recognizes human-understandable concepts such as red color or round shape, and a simple predictor then maps those concepts to a final class. The difficulty is that performance and interpretability pull in opposite directions. Variants that replace concept probabilities with high-dimensional concept embeddings, such as the Concept Embedding Model of Zarlenga and colleagues, boost accuracy but destroy interpretability, because the individual dimensions of an embedding carry no semantic meaning. Conversely, architectures that embed logical rules into the predictor improve transparency but typically learn those rules from data alone, leaving no room for human domain knowledge, and they come with no theoretical guarantee linking overall performance to the quality of individual modules.

The new architecture addresses both gaps with two modules working in tandem. The first is an attribute recognition model, a neural network that inspects an input image and outputs probability vectors for a set of high-level categorical attributes. Instead of treating every binary concept independently, NPCs group concepts into attributes: color, shape, symbol, and text, for example, in the case of traffic signs. Each probability vector expresses the likelihood of the various values that attribute can take. The second module is a task predictor built on a probabilistic circuit, a type of tractable graphical model represented as a rooted directed acyclic graph whose internal nodes are weighted sums and products. Because such circuits satisfy structural properties called smoothness and decomposability, any joint, marginal, or conditional probability over attributes and classes can be computed in at most two passes through the graph, in time linear in its size. The circuit embeds logical rules, either learned implicitly from data or supplied explicitly by humans, within its very structure and parameters.

Formally, the model rests on two mild assumptions. The first, sufficient attributes, states that the class label is conditionally independent of the raw input once the attributes are known; for a traffic sign, knowing the shape, color, symbol, and text is enough to identify the class without consulting the pixels again. The second, complete information, states that the attributes are mutually independent given the input, since the image fully encodes them. Under these assumptions, the probability that an image belongs to a class is computed by summing, over every combination of attribute values, the product of the likelihood of that combination and the circuit’s conditional probability of the class given those attributes. The predicted class is simply the one with the highest score. This compositional formula is what makes the system interpretable: every term in the sum corresponds to a human-readable attribute assignment and an explicit rule linking attributes to classes.

Training proceeds in three stages. First, the attribute recognition model is trained in a multi-task learning framework, with one task per attribute and cross-entropy losses weighted by the size of each attribute’s value space so that no single attribute dominates. Second, the probabilistic circuit is constructed in one of two ways. In the data-driven approach, the LearnSPN algorithm recursively clusters variables and data to learn a circuit structure, and maximum likelihood estimation via the CCCP algorithm, which converges monotonically, optimizes its weights. In the knowledge-injected approach, humans supply weighted logical rules, such as the conjunction of circle, red, and slash indicating a no-entry sign, and the circuit is built as a two-layer sum-of-products formula whose edge weights equal the frequency of each rule. The authors prove that this manually constructed circuit exactly models the empirical joint distribution over attributes and classes. In the third stage, the two modules are fine-tuned jointly end to end, using stochastic gradient descent for the neural network and projected gradient descent to keep the circuit weights positive.

A central theoretical contribution is a compositional error bound, which the authors believe is the first performance guarantee of its kind for compositional bottleneck models. Under the two assumptions, the total error of an NPC, measured as the expected total variation distance between the learned and true class distributions, is upper-bounded by the sum of the errors of the individual attribute recognizers plus twice the error of the probabilistic circuit. In plain terms, improving the recognition of any single attribute reliably improves the whole system, and the overall error can be traced to specific, measurable components. This stands in sharp contrast to end-to-end networks, where a change in one layer can have unpredictable effects elsewhere, and it gives practitioners a principled way to decide where to invest in better modules.

Interpretability is delivered through two complementary explanation types. The Most Probable Explanation identifies the attribute assignment that contributes most to a predicted class, answering the question of why the model chose what it did. The authors introduce an alignment rate, the proportion of correct predictions whose most probable explanation matches the ground-truth attributes, as a reliability metric; across most of their experiments this rate approaches one hundred percent, indicating the model reasons much as a human would. Counterfactual Explanations work in the opposite direction: when the model errs, a projected gradient ascent procedure searches for minimal adjustments to the attribute probabilities that would flip the prediction to the correct class. In one striking example, the model misread the text on a traffic sign as 120 and predicted a speed limit of 120; the counterfactual showed that reading the text as 80, even with only about fifty percent confidence, would have produced the correct classification, confirming that the model’s arithmetic and reasoning chain were sound and only one attribute was at fault.

Empirically, the team evaluated NPCs on four benchmarks: MNIST-Addition, where two digit images must be summed; the German Traffic Sign Recognition Benchmark with 39,209 images; CelebA, a large face dataset; and AwA2, containing 37,322 animal images. NPCs outperformed representative concept-based baselines, including the original concept bottleneck model, its hybrid variant, the Concept Embedding Model, and the Deep Concept Reasoner, and remained competitive with, and on two datasets superior to, fully end-to-end black-box networks, all without a single uninterpretable component. Ablation studies showed that grouping concepts into attributes preserves relational constraints and improves recognition accuracy, that removing decisive attributes such as symbol or text severely degrades performance while removing non-decisive ones like color barely matters, and that data-driven circuits excel on complex datasets while knowledge-injected circuits suffice for simpler ones. Human interventions on attribute values also improved accuracy in line with expectations, and when all attributes were corrected the models reached perfect accuracy on the two simpler benchmarks.

The authors are candid about limitations. The attribute recognition network itself remains a black box vulnerable to spurious correlations, the conditional independence assumption can introduce reasoning shortcuts when confounders are present, and the computational cost of enumerating attribute combinations grows exponentially with the number of attributes. They sketch remedies ranging from large language model generated concept annotations to tensorized circuit implementations such as Einsum Networks. Even so, the message of the work is clear: the long-assumed trade-off between accuracy and transparency is not an iron law. By composing neural perception with tractable probabilistic reasoning, Neural Probabilistic Circuits show that machines can be made to show their work, a capability that may prove essential wherever automated decisions must be trusted, audited, and defended.

Subject of Research: An interpretable machine learning architecture combining neural attribute recognition with probabilistic circuits for logical reasoning

Article Title: Neural Probabilistic Circuits: Enabling Compositional and Interpretable Predictions Through Logical Reasoning

Article References: Chen, W., Yu, S., Shao, H., Sha, L., & Zhao, H. (2026). Neural Probabilistic Circuits: Enabling Compositional and Interpretable Predictions Through Logical Reasoning. Machine Learning, 115(9), Article 207. https://doi.org/10.1007/s10994-026-07118-7

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07118-7

Keywords: machine learning, interpretability, probabilistic circuits, concept bottleneck models, neuro-symbolic AI, logical reasoning, explainable AI, deep learning, image classification, counterfactual explanations, error bounds, graphical models