Causal Neural Probabilistic Circuits

Han Zhao; Weixin Chen

REVIEW 3 major objections 3 minor 58 references

Causal Neural Probabilistic Circuits

T0 review · 3 major / 3 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper argues that compiling the causal graph over concepts into a probabilistic circuit, then fusing its interventional marginals with a neural predictor via a product of experts, makes expert interventions far more effective in concept

desk verdict CNPC is a sensible new model for intervention propagation in CBMs, but the headline OOD gains look more like a fixed-alpha prior effect than evidence for causal structure. read the letter →

arxiv 2603.01372 v2 pith:FYSW5HAO submitted 2026-03-02 cs.LG cs.AI

Weixin Chen , Han Zhao This is my paper · ORCID

classification cs.LGcs.AI

keywords conceptbottleneckmodelsprobabilisticcircuitscausalinferenceinterventionsproductofexpertsinterpretabilityout-of-distribution

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks how a concept-bottleneck model should respond when a human expert corrects one attribute prediction at test time. Standard CBMs overwrite the corrected attribute and leave the others untouched, ignoring causal dependencies among attributes. The paper proposes CNPC, which compiles a causal graph over attributes and the class label into a probabilistic circuit that supports exact interventional inference, then fuses the circuit's interventional marginals with the neural predictor's distribution via a product of experts. The central claim is that this causal propagation substantially improves task accuracy under intervention compared with five baselines, especially when the neural predictor is unreliable under distribution shift. The paper also proves a compositional error bound identifying conditions under which this fusion improves over non-causal propagation.

What carries the argument

The central object is a causal probabilistic circuit: a sum-product network compiled from the causal graph over attributes and the class label via variable elimination, so that interventional queries are evaluated by clamping the intervened variable's conditional-probability entries to 1 and running one forward pass. It carries the argument by exactly computing interventional marginals, giving the causal propagation that ordinary PCs and standard CBMs lack. The second piece is a product-of-experts (PoE) fusion—the normalized geometric mean of the clamped neural predictive distribution and the causal PC's interventional marginal, with weight α—which supplies the input-conditioned evidence tha

What would settle it

Take a dataset with a known true causal graph and strong attribute dependencies, train CNPC with the correct graph, then retrain it with a deliberately corrupted graph (reverse an edge or add a spurious edge). If intervention accuracy does not drop relative to the correct-graph version, the causal-propagation mechanism is not actually carrying the reported gains. Conversely, if a version with α=0 (no causal PC fusion) matches CNPC's accuracy, the causal component is superfluous.

Watch

Extended reading notes

Core claim

CNPC claims that exactly modeling the interventional class distribution—P(Y|X) after setting an attribute to its ground-truth value—is hard with standard CBM modules because the interventional conditional over attributes cannot be derived from either the neural predictor or a non-causal probabilistic circuit. The paper's model replaces the label-predictor PC with a causal PC compiled from the causal graph, which computes interventional marginals that respect causal dependencies by construction. Since the exact interventional conditional is still unavailable, CNPC approximates it as a product of experts: the geometric mean of the clamped neural predictive distribution (which preserves input e

Load-bearing premise

The causal graph over attributes and the class label is assumed known and correct (with no unobserved confounders), so if the graph is misspecified, the interventional marginals that the whole method leans on will propagate wrong dependencies.

Editorial extensions

If this is right

Under intervention on one attribute, CNPC's update also shifts predictions for causally downstream attributes, so task accuracy improves without additional expert corrections.
In OOD settings where the attribute predictor is unreliable, the causal PC's interventional marginals become the dominant source (α=0.9), so CNPC retains high intervention efficiency.
The compositional error bounds imply that improving either the attribute predictor or the causal PC's interventional marginals tightens the overall interventional error, giving a modular roadmap for building better CBMs.
CNPC reduces to NPC when α=0, so the causal graph and PoE are the only additions; ablations show intermediate α beats both extremes in benign settings, suggesting causal propagation is not a substitute for input evidence.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable extension: tie α to an OOD detector score so that inputs flagged as out-of-distribution lean more heavily on the causal circuit's interventional marginals—this would make the fusion adaptive rather than a fixed global weight.
The depth-based intervention ordering is a heuristic; near-deterministic parents of the class variable could yield larger gains per intervention, so a principled selection policy remains an open problem the paper itself acknowledges.
Because the causal PC supports counterfactual inference, the same machinery could answer 'what if a different attribute value had been set'—a class of questions standard CBMs cannot address.
The whole advantage hinges on a correct causal graph; whether the method degrades gracefully under graph misspecification is untested, and a version that learns the graph jointly would broaden applicability.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things to know. First, CNPC is a legitimate new model: a neural attribute predictor plus a causal PC, fused with a product-of-experts rule at intervention time. The math is correct and the paper is readable. Second, the headline result—large OOD gains—is not actually shown to come from causal structure. The authors fix alpha=0.9 in all OOD experiments without a held-out OOD split, and at that weight the PoE is dominated by the causal PC's interventional marginal, which doesn't look at the image. That could help under distribution shift even if the graph were wrong, because it acts as a robust prior. The paper never tests graph misspecification or compares against a non-causal fixed prior, so the causal mechanism claim is unproven.

What's good: The combination is new relative to NPC, SCBM, and C2BM. The causal PC compilation is standard, but using it for CBM intervention propagation with a PoE is a sensible idea. The theoretical bounds in Corollaries 3 and 4 are correct KL decompositions; they just aren't tight—equality requires the modules to be essentially perfect. The experiments cover five datasets and three OOD shifts, which is a broad evaluation. The paper is honest about its limitations, including adaptive alpha and graph availability.

Soft spots: (1) The alpha choice is the big one. Setting alpha=0.9 without OOD validation, then reporting the test-set results, risks overfitting to the test set. Their own ablation shows the optimal alpha varies by setting (0.7 to 0.9), so 0.9 isn't clearly the right number. (2) No error bars on any figure. Averaging over three seeds is fine, but without variance we can't assess the significance of the claimed 23-24% gaps. (3) The CelebA graph is manually annotated, and there's no sensitivity analysis to graph errors. (4) Code isn't public. (5) The abstract's phrase 'inherently respects causal dependencies' is true about the circuit, but the empirical support for causal propagation is indirect.

Who this is for: researchers working on concept bottleneck models, test-time interventions, or tractable causal inference. They'll get a clear method and a useful baseline. But they should treat the OOD claims as preliminary until the alpha selection and graph sensitivity are settled.

Recommendation: send to peer review. The paper deserves referee time; the core idea is solid and the problems are fixable with additional experiments, not fundamental. I'd ask for an ablation that isolates whether the causal graph matters—e.g., compare alpha=0.9 with a wrong graph, and with a non-causal PC that just uses the observational marginal.

Referee Report

3 major / 3 minor

Summary. The paper proposes Causal Neural Probabilistic Circuits (CNPC), a concept-bottleneck architecture that couples a neural attribute predictor with a probabilistic circuit compiled from a known causal graph over attributes and the class label. Under attribute interventions, CNPC approximates the interventional attribute posterior by a product of experts (PoE) between the clamped neural predictor and the causal PC's interventional marginal, then computes the class posterior from the causal PC's conditional. The authors provide KL-divergence bounds for NPC and CNPC under interventions, and evaluate on Asia, Sachs, MNISTAdd, cMNISTAdd, and CelebA in benign and out-of-distribution settings, reporting higher task accuracy than five baselines across intervention counts, especially under OOD shifts.

Significance. If the empirical claims hold, CNPC is a practically useful way to propagate expert interventions through causal structure in concept bottleneck models, with exact and tractable probabilistic causal inference provided by the compiled circuit. The paper has strengths: the compilation of a causal graph into a PC is standard but the application to CBM interventions is sensible; the theoretical bounds are correct as bounds; the experiments cover several OOD settings and the code is promised in the supplementary material. However, the central causal claim is not yet isolated from the choice of the PoE weight α, and the theoretical 'conditions' under which CNPC matches the ground truth are equality conditions that are only generically satisfied when the modules are essentially perfect. The paper's contribution is therefore plausible and potentially valuable, but the current evidence overstates the role of correct causal propagation.

major comments (3)

[Sec. 6.1, Eq. (4), Fig. 5] The OOD headline is obtained with α=0.9 fixed across all datasets and intervention counts, with no held-out OOD validation (Sec. 6.1). At α=0.9, the PoE in Eq. (4) is dominated by the input-independent causal-PC marginal P_do_w(A), so the reported OOD gains may be produced by a robust fixed prior rather than by correct causal propagation. Figure 5 shows that OOD task accuracy is strongly α-dependent, so the comparison to baselines depends on this choice. The authors should select α on an OOD validation split, or at least report results across a range of α, and include an ablation that replaces the causal PC with an input-independent non-causal marginal to separate the causal mechanism from the robust-prior effect.
[Sec. 5, Corollaries 3 and 4, Appendix F] The theoretical claim that CNPC 'matches the ground-truth interventional class distribution' under certain conditions is not supported by the stated conditions. In the proof of Corollary 7 (restating Corollary 4), equality in the PoE KL decomposition requires P_theta_doj(a|x)=P_w_doj(a) for all a, i.e., an input-dependent distribution must equal an input-independent distribution for every x. This is a very restrictive, essentially perfect-module condition, and the same applies to the ratio-constant equality condition in Corollary 3. The bounds are valid, but they do not characterize when the actual error is small, and the comparison B_CNPC≤B_NPC is a comparison of upper bounds, not of true errors. The text in Section 5 and the abstract should be revised to state this limitation explicitly.
[Sec. 3.1, Assumption 2; Appendix E.1.1, Fig. 7b] The method requires the causal graph over attributes and the class label to be known and correct, but the paper provides no sensitivity analysis to graph misspecification. In particular, the CelebA graph is manually annotated (Appendix E.1.1) and is not validated. If the graph is misspecified, the causal PC's interventional marginals are incorrect, and the PoE will propagate wrong dependencies. Because α=0.9 places large weight on these marginals, the reported OOD gains could survive even with an incorrect graph, which would weaken the causal interpretation. An ablation with a deliberately misspecified or shuffled graph, or with a graph learned from data, would directly address this concern.

minor comments (3)

[Eq. (3) and Appendix D] In Eq. (3), P(Y|A_{1:K}) is written without the intervention superscript. The surrounding text says this follows from invariance, but the notation is confusing because the intervening distribution is otherwise denoted with do. It would be clearer to write P_doj(Y|A)=P(Y|A) explicitly. Additionally, the step 'uses Assumption 1' in Appendix D should explicitly state that the conditional independence Y⊥X|A is assumed to be stable under interventions (or justify it from d-separation if X is included in the causal graph).
[Figures 3, 4, 6] All results are averaged over three random seeds, but no error bars or standard deviations are shown. Since the claimed OOD improvements are substantial, this is not fatal, but adding error bars or reporting variance in the text would increase confidence.
[Sec. 6.1] The choice α=0.9 in OOD settings is stated as a fixed value 'across all datasets and all intervened attributes', but the ablation study later shows that the optimal α differs across OOD types (e.g., α=0.7 for adversarial MNISTAdd, α=0.9 for CelebA). This inconsistency should be discussed more prominently; it further motivates an adaptive α or a principled selection rule.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CNPC derivation is self-contained and the theory is proved in the appendix.

full rationale

CNPC's derivation chain is not circular. The attribute predictor and causal PC are separate, independently defined modules; their fusion is an explicit geometric-mean/PoE choice in Eq. (4), and the class prediction in Eq. (5) is an exact summation over the PC conditional. The theoretical bounds (Theorem 5, Corollaries 3-4) are proved in Appendix F using a standard three-step KL/log-sum/Hoelder decomposition; although Theorem 5 is described as analogous to the authors' prior NPC result, it is fully proven here rather than assumed. The comparison B_CNPC <= B_NPC when KL(P*_do(A|X)||P_do_w(A)) <= KL(P*_do(A|X)||P_do_theta(A|X)) follows directly from the stated bounds and is a conditional statement, not an assertion that the PoE equals the ground-truth interventional conditional. Empirically, CNPC is compared against five external baselines on standard benchmarks; the OOD evaluation fixes alpha=0.9 without a held-out OOD validation split (Sec. 6.1), which limits the strength of the OOD claim, but this is a hyperparameter-selection/limitation issue rather than a circular reduction. No quantity used as a 'prediction' is a renamed fitted value, and no load-bearing argument relies on a self-citation in place of proof. The self-citations to NPC [12,13,37] supply background, baselines, and a theorem that is reproved, so they do not create circularity.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

CNPC introduces no new postulated entities; it relies on existing tools (causal probabilistic circuits, product of experts) and on the causal-graph assumption that is standard in the CBM-with-causality literature. The main free parameter is α, which is hand-set in OOD settings.

free parameters (1)

α (PoE weight) = 0.9 in OOD settings; selected on validation set in benign settings (e.g., peak at 0.3)
Controls the trade-off between the neural attribute predictor and the causal PC's interventional marginal. In OOD settings it is fixed at 0.9 across all datasets and intervention counts because OOD validation data is unavailable; the ablation in Section 6.3 shows performance is sensitive to α.

assumptions (6)

domain assumption Assumption 1 (Sufficient attributes): Y ⊥ X | A_{1:K}
Section 3.1; the entire CBM framework and the derivation of Eq. (3) depend on the class label being conditionally independent of the input given the attributes.
domain assumption Assumption 2: Pa_C(Y) ⊆ {A_{1:K}} ⊆ NDC(Y)
Section 3.1; ensures attributes can influence the class but the class does not influence attributes. Used in the Appendix D derivation of the interventional class distribution.
domain assumption No unobserved confounders among the variables in C
Section 3.1; required for the causal PC's interventional marginals to be valid and for causal inference from the compiled circuit to be sound.
domain assumption The causal graph structure C is available and correct
Section 3.1. In the experiments, the CelebA graph is manually annotated (Appendix E.1.1) and the MNISTAdd graph is synthetic; misspecification would break the propagation logic.
standard math Causal Markov property and invariance of causal mechanisms under interventions
Used in Appendix D to derive Eq. (3); standard assumptions in causal inference (Pearl 2009).
domain assumption The causal PC's parameters, estimated by maximum likelihood, are accurate
Section 4.2; the interventional marginal P^do_w(A) is computed from the fitted PC, and if the fit is poor, the PoE receives a bad prior.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Causal Neural Probabilistic Circuits." pith.science (2026). https://pith.science/paper/FYSW5HAO

@misc{pith2026260301372,
  author       = {Pith},
  title        = {Pith review of: Causal Neural Probabilistic Circuits},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FYSW5HAO}},
  note         = {Machine review of arXiv:2603.01372}
}

read the original abstract

Concept Bottleneck Models (CBMs) enhance the interpretability of end-to-end neural networks by introducing a layer of concepts and predicting the class label from the concept predictions. A key property of CBMs is that they support interventions, i.e., domain experts can correct mispredicted concept values at test time to improve the final accuracy. However, typical CBMs apply interventions by overwriting only the corrected concept while leaving other concept predictions unchanged, which ignores causal dependencies among concepts. To address this, we propose the Causal Neural Probabilistic Circuit (CNPC), which combines a neural attribute predictor with a causal probabilistic circuit compiled from a causal graph. This circuit supports exact, tractable causal inference that inherently respects causal dependencies. Under interventions, CNPC models the class distribution based on a Product of Experts (PoE) that fuses the attribute predictor's predictive distribution with the interventional marginals computed by the circuit. We theoretically characterize the compositional interventional error of CNPC w.r.t. its modules and identify conditions under which CNPC closely matches the ground-truth interventional class distribution. Experiments on five benchmark datasets in both in-distribution and out-of-distribution settings show that, compared with five baseline models, CNPC achieves higher task accuracy across different numbers of intervened attributes.

Figures

Figures reproduced from arXiv: 2603.01372 by the authors.

**Figure 1.** Model architectures and intervention procedures of a CBM and a CNPC. A typical CBM (top-left module + top-right module) performs the intervention on A1 by replacing the neural network’s predictions for A1 with the ground-truth distribution, while leaving the predictions for other attributes unchanged. In contrast, CNPC (top-left module + bottom module) combines a neural network with a causal PC compiled from a causa… view at source ↗

**Figure 2.** A causal PC that compiles the causal graph V2 ← V1 → V3. An arithmetic circuit is then constructed by establishing a sum node for the sum-out operator while establishing a product node for the multiplication operator. For simplicity, suppose all variables are binary5 , and we use vi and v¯i to denote Vi = 1 and Vi = 0. The resulting circuit is illustrated in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Task accuracy of all models in the benign setting on the Asia, Sachs, MNISTAdd, and CelebA datasets under varying numbers of attribute interventions. All results are averaged across three random seeds [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: Task accuracy of all models in OOD settings on the MNISTAdd and CelebA datasets under varying numbers of attribute interventions. All results are averaged across three random seeds. in the causal graph, i.e., the longest directed path from the root nodes. For CNPC, the…

**Figure 5.** Figure 5: Task accuracy of CNPC in both benign and OOD settings with α varying from 0.0 to 1.0 in increments of 0.1. Left: Performance on MNISTAdd with one intervened attribute. Middle/Right: Performance on CelebA with two/four intervened attributes. All results are averaged acr…

**Figure 6.** Figure 6: Attribute accuracy (left) and task accuracy (right) of all models on cMNISTAdd. Spurious-correlation shifts [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 7.** Figure 7: Causal graphs over attributes A1:K and the class label Y . Left: The causal graph for MNISTAdd, with attributes A1: the first digit and A2: the second digit. Right: The causal graph for CelebA, with attributes A1: Attractive, A2: Mouth_Slightly_Open, A3: Smiling, A4: W…

**Figure 8.** Figure 8: Instances from the cMNISTAdd dataset. Top: Examples from the training set. Bottom: Examples from the test set. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]

**Figure 9.** Figure 9: Attribute accuracy of CNPC and baseline models in the benign setting on the Asia, Sachs, [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]

**Figure 10.** Figure 10: Attribute accuracy of CNPC and baseline models in OOD settings on the MNISTAdd and [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 3 linked inside Pith

[1]

Imagenet: A large-scale hierarchical image database

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. InIEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255. IEEE Computer Society, 2009

2009
[2]

Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In 7th International Conference on Learning Representations (ICLR). OpenReview.net, 2019

2019
[3]

Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nat

Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nat. Mach. Intell., 1(5):206–215, 2019

2019
[4]

Concept bottleneck models

Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. InProceedings of the 37th International Conference on Machine Learning (ICML), volume 119 ofProceedings of Machine Learning Research, pages 5338–5348. PMLR, 2020

2020
[5]

Right for the right concept: Revising neuro-symbolic concepts by interacting with their explanations

Wolfgang Stammer, Patrick Schramowski, and Kristian Kersting. Right for the right concept: Revising neuro-symbolic concepts by interacting with their explanations. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3619–3629. Computer Vision Foundation / IEEE, 2021

2021
[6]

Novoa, Justin Ko, Susan M

Andre Esteva, Brett Kuprel, Roberto A. Novoa, Justin Ko, Susan M. Swetter, Helen M. Blau, and Sebastian Thrun. Dermatologist-level classification of skin cancer with deep neural networks. Nat., 542(7639):115–118, 2017

2017
[7]

Clinically applicable deep learning for diagnosis and referral in retinal disease.Nature medicine, 24(9):1342–1350, 2018

Jeffrey De Fauw, Joseph R Ledsam, Bernardino Romera-Paredes, Stanislav Nikolov, Nenad Tomasev, Sam Blackwell, Harry Askham, Xavier Glorot, Brendan O’Donoghue, Daniel Visentin, et al. Clinically applicable deep learning for diagnosis and referral in retinal disease.Nature medicine, 24(9):1342–1350, 2018

2018
[8]

Concept embedding models: Beyond the accuracy- explainability trade-off

Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frédéric Precioso, Stefano Melacci, Adrian Weller, Pietro Lió, and Mateja Jamnik. Concept embedding models: Beyond the accuracy- explainability trade-off. InAdvances in Neural Information Processing Systems 35: Annual Con...

2022

Show all 58 references

[9]

Oikarinen, Subhro Das, Lam M

Tuomas P. Oikarinen, Subhro Das, Lam M. Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. InThe Eleventh International Conference on Learning Representations (ICLR). OpenReview.net, 2023

2023
[10]

Causal inference in statistics: An overview

Judea Pearl. Causal inference in statistics: An overview. 2009

2009
[11]

The MIT press, 2017

Jonas Peters, Dominik Janzing, and Bernhard Schölkopf.Elements of causal inference: founda- tions and learning algorithms. The MIT press, 2017

2017
[12]

Neural probabilistic circuits: Enabling compositional and interpretable predictions through logical reasoning.arXiv preprint arXiv:2501.07021, 2025

Weixin Chen, Simon Yu, Huajie Shao, Lui Sha, and Han Zhao. Neural probabilistic circuits: Enabling compositional and interpretable predictions through logical reasoning.arXiv preprint arXiv:2501.07021, 2025

2025 arXiv
[13]

Neural probabilistic circuits: An overview

Weixin Chen, Simon Yu, Huajie Shao, Lui Sha, and Han Zhao. Neural probabilistic circuits: An overview. InEighth Workshop on Tractable Probabilistic Modeling, 2025

2025
[14]

Domingos

Hoifung Poon and Pedro M. Domingos. Sum-product networks: A new deep architecture. In Proceedings of the Twenty-Seventh Conference on Uncertainty in Artificial Intelligence (UAI), pages 337–346. AUAI Press, 2011

2011
[15]

Probabilistic circuits: A unifying framework for tractable probabilistic models.UCLA

Y Choi, Antonio Vergari, and Guy Van den Broeck. Probabilistic circuits: A unifying framework for tractable probabilistic models.UCLA. URL: http://starai. cs. ucla. edu/papers/ProbCirc20. pdf, page 6, 2020

2020
[16]

Probabilistic circuits: Representations, inference, learning and applications.AAAI Tutorial, 2020

Antonio Vergari, YooJung Choi, Robert Peharz, and Guy Van den Broeck. Probabilistic circuits: Representations, inference, learning and applications.AAAI Tutorial, 2020. 10

2020
[17]

Causal inference using tractable circuits.arXiv preprint arXiv:2202.02891, 2022

Adnan Darwiche. Causal inference using tractable circuits.arXiv preprint arXiv:2202.02891, 2022

2022 arXiv
[18]

Training products of experts by minimizing contrastive divergence.Neural computation, 14(8):1771–1800, 2002

Geoffrey E Hinton. Training products of experts by minimizing contrastive divergence.Neural computation, 14(8):1771–1800, 2002

2002
[19]

Dietterich

Dan Hendrycks and Thomas G. Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In7th International Conference on Learning Representations (ICLR). OpenReview.net, 2019

2019
[20]

Goodfellow, Jonathon Shlens, and Christian Szegedy

Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing ad- versarial examples. In3rd International Conference on Learning Representations (ICLR), 2015

2015
[22]

Moritz Vandenhirtz, Sonia Laguna, Ricards Marcinkevics, and Julia E. V ogt. Stochastic concept bottleneck models. InAdvances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems (NeurIPS), 2024

2024
[23]

Addressing leakage in concept bottle- neck models

Marton Havasi, Sonali Parbhoo, and Finale Doshi-Velez. Addressing leakage in concept bottle- neck models. InAdvances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems (NeurIPS), 2022

2022
[24]

Causally reliable concept bottleneck models

Giovanni de Felice, Arianna Casanova Flores, Francesco De Santis, Silvia Santini, Johannes Schneider, Pietro Barbiero, and Alberto Termine. Causally reliable concept bottleneck models. InAdvances in Neural Information Processing Systems 39: Annual Conference on Neural Informat...

2025
[25]

Causal concept graph models: Beyond causal opacity in deep learning

Gabriele Dominici, Pietro Barbiero, Mateo Espinosa Zarlenga, Alberto Termine, Martin Gjoreski, Giuseppe Marra, and Marc Langheinrich. Causal concept graph models: Beyond causal opacity in deep learning. InThe Thirteenth International Conference on Learning Representations (ICL...

2025
[26]

A closer look at the intervention procedure of concept bottleneck models

Sungbin Shin, Yohan Jo, Sungsoo Ahn, and Namhoon Lee. A closer look at the intervention procedure of concept bottleneck models. InInternational Conference on Machine Learning (ICML), volume 202 ofProceedings of Machine Learning Research, pages 31504–31520. PMLR, 2023

2023
[27]

Learning from uncertain concepts via test time interventions

Ivaxi Sheth, Aamer Abdul Rahman, Laya Rafiee Sevyeri, Mohammad Havaei, and Samira Ebrahimi Kahou. Learning from uncertain concepts via test time interventions. In Workshop on trustworthy and socially responsible machine learning, NeurIPS, 2022

2022
[28]

Interactive concept bottleneck models

Kushal Chauhan, Rishabh Tiwari, Jan Freyberg, Pradeep Shenoy, and Krishnamurthy Dvi- jotham. Interactive concept bottleneck models. InThirty-Seventh AAAI Conference on Artificial Intelligence (AAAI), pages 5948–5955. AAAI Press, 2023

2023
[29]

Learning to intervene on concept bottlenecks

David Steinmann, Wolfgang Stammer, Felix Friedrich, and Kristian Kersting. Learning to intervene on concept bottlenecks. InForty-first International Conference on Machine Learning (ICML), volume 235 ofProceedings of Machine Learning Research, pages 46556–46571. PMLR / OpenRevi...

2024
[30]

Post-hoc concept bottleneck models

Mert Yüksekgönül, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. In The Eleventh International Conference on Learning Representations (ICLR). OpenReview.net, 2023

2023
[31]

VLG-CBM: training concept bottleneck models with vision-language guidance

Divyansh Srivastava, Ge Yan, and Lily Weng. VLG-CBM: training concept bottleneck models with vision-language guidance. InAdvances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems (NeurIPS), 2024

2024
[32]

Causal diagrams for empirical research.Biometrika, 82(4):669–688, 1995

Judea Pearl. Causal diagrams for empirical research.Biometrika, 82(4):669–688, 1995. 11

1995
[33]

A general identification condition for causal effects

Jin Tian and Judea Pearl. A general identification condition for causal effects. InProceedings of the Eighteenth National Conference on Artificial Intelligence and Fourteenth Conference on Innovative Applications of Artificial Intelligence (AAAI), pages 567–573. AAAI Press / T...

2002
[34]

An algorithm for fast recovery of sparse causal graphs.Social science computer review, 9(1):62–72, 1991

Peter Spirtes and Clark Glymour. An algorithm for fast recovery of sparse causal graphs.Social science computer review, 9(1):62–72, 1991

1991
[35]

Learning bayesian networks: The combination of knowledge and statistical data.Machine learning, 20(3):197–243, 1995

David Heckerman, Dan Geiger, and David M Chickering. Learning bayesian networks: The combination of knowledge and statistical data.Machine learning, 20(3):197–243, 1995

1995
[36]

Optimal structure identification with greedy search.Journal of machine learning research, 3(Nov):507–554, 2002

David Maxwell Chickering. Optimal structure identification with greedy search.Journal of machine learning research, 3(Nov):507–554, 2002

2002
[37]

Understanding and improving adversarial robustness of neural probabilistic circuits

Weixin Chen and Han Zhao. Understanding and improving adversarial robustness of neural probabilistic circuits. InAdvances in Neural Information Processing Systems 39: Annual Conference on Neural Information Processing Systems (NeurIPS), 2025

2025
[38]

Bucket elimination: A unifying framework for reasoning.Artificial Intelligence, 113(1-2):41–85, 1999

Rina Dechter. Bucket elimination: A unifying framework for reasoning.Artificial Intelligence, 113(1-2):41–85, 1999

1999
[39]

MIT press, 2009

Daphne Koller and Nir Friedman.Probabilistic graphical models: principles and techniques. MIT press, 2009

2009
[40]

Cambridge university press, 2009

Adnan Darwiche.Modeling and reasoning with Bayesian networks. Cambridge university press, 2009

2009
[41]

Compiling bayesian networks using variable elimination

Mark Chavira and Adnan Darwiche. Compiling bayesian networks using variable elimination. InProceedings of the 20th International Joint Conference on Artificial Intelligence (IJCAI), pages 2443–2449, 2007

2007
[42]

A differential approach to inference in bayesian networks.J

Adnan Darwiche. A differential approach to inference in bayesian networks.J. ACM, 50(3): 280–305, 2003

2003
[43]

Nevin Lianwen Zhang and David L. Poole. Exploiting causal independence in bayesian network inference.J. Artif. Intell. Res., 5:301–328, 1996

1996
[44]

Denoising diffusion probabilistic models

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. InAd- vances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems (NeurIPS), 2020

2020
[45]

Generative adversarial networks.Communications of the ACM, 63(11):139–144, 2020

Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks.Communications of the ACM, 63(11):139–144, 2020

2020
[46]

Combining probability distributions: A critique and an annotated bibliography.Statistical Science, 1(1):114–135, 1986

Christian Genest and James V Zidek. Combining probability distributions: A critique and an annotated bibliography.Statistical Science, 1(1):114–135, 1986

1986
[47]

Selecting weighting factors in logarithmic opinion pools

Tom Heskes. Selecting weighting factors in logarithmic opinion pools. InAdvances in Neural Information Processing Systems 10 (NeurIPS), pages 266–272. The MIT Press, 1997

1997
[48]

Steffen L Lauritzen and David J Spiegelhalter. Local computations with probabilities on graphical structures and their application to expert systems.Journal of the Royal Statistical Society: Series B (Methodological), 50(2):157–194, 1988

1988
[49]

Learning bayesian networks with the bnlearn r package.Journal of statistical software, 35:1–22, 2010

Marco Scutari. Learning bayesian networks with the bnlearn r package.Journal of statistical software, 35:1–22, 2010

2010
[50]

Causal protein-signaling networks derived from multiparameter single-cell data.Science, 308(5721): 523–529, 2005

Karen Sachs, Omar Perez, Dana Pe’er, Douglas A Lauffenburger, and Garry P Nolan. Causal protein-signaling networks derived from multiparameter single-cell data.Science, 308(5721): 523–529, 2005. 12

2005
[51]

Deepproblog: Neural probabilistic logic programming

Robin Manhaeve, Sebastijan Dumancic, Angelika Kimmig, Thomas Demeester, and Luc De Raedt. Deepproblog: Neural probabilistic logic programming. InAdvances in Neural Informa- tion Processing Systems 31: Annual Conference on Neural Information Processing Systems, (NeurIPS), pages...

2018
[52]

Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, 2002

Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, 2002

2002
[53]

Invariant risk mini- mization.arXiv preprint arXiv:1907.02893, 2019

Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk mini- mization.arXiv preprint arXiv:1907.02893, 2019

1907 arXiv
[54]

Deep learning face attributes in the wild

Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. InProceedings of the IEEE international conference on computer vision (CVPR), pages 3730–3738, 2015

2015
[55]

Towards deep learning models resistant to adversarial attacks

Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. InICLR, 2018

2018
[56]

A baseline for detecting misclassified and out-of-distribution examples in neural networks

Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In5th International Conference on Learning Representations (ICLR). OpenReview.net, 2017

2017
[57]

Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In6th International Conference on Learning Representations (ICLR). OpenReview.net, 2018

2018
[58]

Domingos

Robert Gens and Pedro M. Domingos. Learning the structure of sum-product networks. In ICML, pages 873–880, 2013

2013
[59]

What would the outcome have been had the attribute been set to a different value?

Han Zhao, Pascal Poupart, and Geoffrey J. Gordon. A unified approach for learning the parameters of sum-product networks. InNeurIPS, pages 433–441, 2016. 13 A Discussion In this section, we discuss the limitations of the proposed method and outline potential solutions, which a...

2016

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

Imagenet: A large-scale hierarchical image database

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. InIEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255. IEEE Computer Society, 2009

2009

[2] [2]

Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In 7th International Conference on Learning Representations (ICLR). OpenReview.net, 2019

2019

[3] [3]

Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nat

Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nat. Mach. Intell., 1(5):206–215, 2019

2019

[4] [4]

Concept bottleneck models

Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. InProceedings of the 37th International Conference on Machine Learning (ICML), volume 119 ofProceedings of Machine Learning Research, pages 5338–5348. PMLR, 2020

2020

[5] [5]

Right for the right concept: Revising neuro-symbolic concepts by interacting with their explanations

Wolfgang Stammer, Patrick Schramowski, and Kristian Kersting. Right for the right concept: Revising neuro-symbolic concepts by interacting with their explanations. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3619–3629. Computer Vision Foundation / IEEE, 2021

2021

[6] [6]

Novoa, Justin Ko, Susan M

Andre Esteva, Brett Kuprel, Roberto A. Novoa, Justin Ko, Susan M. Swetter, Helen M. Blau, and Sebastian Thrun. Dermatologist-level classification of skin cancer with deep neural networks. Nat., 542(7639):115–118, 2017

2017

[7] [7]

Clinically applicable deep learning for diagnosis and referral in retinal disease.Nature medicine, 24(9):1342–1350, 2018

Jeffrey De Fauw, Joseph R Ledsam, Bernardino Romera-Paredes, Stanislav Nikolov, Nenad Tomasev, Sam Blackwell, Harry Askham, Xavier Glorot, Brendan O’Donoghue, Daniel Visentin, et al. Clinically applicable deep learning for diagnosis and referral in retinal disease.Nature medicine, 24(9):1342–1350, 2018

2018

[8] [8]

Concept embedding models: Beyond the accuracy- explainability trade-off

Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frédéric Precioso, Stefano Melacci, Adrian Weller, Pietro Lió, and Mateja Jamnik. Concept embedding models: Beyond the accuracy- explainability trade-off. InAdvances in Neural Information Processing Systems 35: Annual Con...

2022