REVIEW 4 major objections 5 minor 43 references
Learning Nonlinear Causal Reductions to Explain Reinforcement Learning Policies
T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read This paper claims that RL policy success and failure can be explained by a single learned high-level cause, and proves this explanation is unique up to rescaling within a class of nonlinear additive-noise causal models.
desk verdict A genuine nonlinear identifiability result, let down by an empirical section that doesn't verify the theorem's assumptions—worth refereeing with major revisions. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The argument is carried by the exact-transformation condition of Definition 2.2: the push-forward of every intervened low-level distribution through $\tau$ must equal the high-level distribution under the corresponding intervention, $\tau_\#[P_L^{(i)}] = P_H^{(\omega(i))}$. The proof of uniqueness (Prop. 4.1) works in the Fourier domain—exact equality of expectations for all shift interventions produces a convolution equation $p_1 \ast (\delta\tau_1 \circ g) = 0$, and the non-vanishing Fourier transform of the noise density leaves only $\delta\tau_1 \equiv 0$. On the practical side, the learning machinery is a consistency loss between push-forward and high-level distributions, regularized by the 1-Wasserstein distance of the standardized high-level cause to the standard normal, over an interpretable function class of Gaussian kernels placed per feature and time step.
What would settle it
Decisive check: on a synthetic additive-noise SCM satisfying the three conditions of Proposition 4.1, train nTCR from many random initializations until the consistency loss reaches machine precision; if any two runs recover high-level causes that differ by more than an additive and multiplicative rescaling, the uniqueness claim is false. A complementary check applies the same restarts to the RL datasets and inspects whether the learned $\tau$ maps converge to a single affine class.
Extended reading notes
Core claim
The central claim is that causal explanations of an RL policy can be computed by learning a nonlinear reduction from the low-level episode model to a high-level model with one cause and one target, and that within a well-defined model class the reduction is identifiable. Concretely, Proposition 4.1 states: if the low-level model is an additive-noise SCM $X_{\pi(1)} = f_1(X_{\pi(1)}) + U_{\pi(1)} + i_{\pi(1)}$, $X_{\pi(0)} = f_0(X_{\pi(0)}, X_{\pi(1)}) + U_{\pi(0)}$ with $U_{\pi(1)}$ independent of $U_{\pi(0)}$, the density of $P_1$ has non-vanishing Fourier transform, and the high-level causal effect is nonzero, then any constructive transformation that is exact for all shift interventions is unique up to additive and multiplicative constants. The paper also proves existence for a family of such models, giving closed-form reduction maps (Prop. 4.2), and validates the framework by recovering ground-truth solutions on synthetic data and by learning interpretable temporal heatmaps on two RL tasks.
Load-bearing premise
The load-bearing premise is that an RL episode is governed by an additive-noise causal model with independent exogenous noises and a noise density whose Fourier transform is nowhere zero, which is what lets the proof 'undo' the convolution and conclude that two exactly consistent reductions must be the same; the paper does not verify this structural condition on the real RL environments.
Editorial extensions
If this is right
- When an exact reduction exists, the learned high-level cause is identified up to additive and multiplicative constants, so the explanation is unambiguous in the limit of observing all shift interventions.
- The method surfaces behavioral patterns that aggregate reward metrics hide, including a clockwise-versus-counterclockwise bias in the pendulum policy and ball-trajectory-dependent difficulties in table tennis.
- The Gaussian-kernel function class keeps the reduction interpretable: the learned weights directly tell which feature, at which time step, contributes to the high-level cause.
- The normality regularization prevents the nonlinear explainer from overfitting the consistency loss with multi-modal high-level causes, and is most beneficial when interventions are weak relative to exogenous noise.
- Because the framework uses only shift interventions on continuous actions, it transfers to other continuous-action control tasks, and the authors argue it can be adapted to discrete action spaces by perturbing the underlying policy parameters.
Reading between the lines
- Beyond the paper: if the uniqueness result could be extended from 'all shift interventions' to the finite intervention sets used in practice (the paper states this as an open conjecture), nTCR explanations would be reproducible across training runs; a direct test is to measure how the consistency loss and the recovered $\tau$ maps vary across restarts on the two RL datasets.
- Beyond the paper: because the interpretable basis is a sum over features and time steps, the learned maps implicitly encode timing; a natural extension is to replace the fixed Gaussian kernels with smooth temporal features such as derivatives or lagged values to test hypotheses about when, not just which, actions matter.
- Beyond the paper: the deterministic Pendulum environment is a stress test for the additive-noise assumption; adding controllable exogenous noise to the simulator and checking whether the recovered reductions change in the direction predicted by the theory would separate the method's robustness from its assumptions.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper extends Targeted Causal Reduction (TCR) to nonlinear settings (nTCR) and applies it to explain trained RL policies. The low-level model is an SCM of episode variables; shift interventions on actions are used to probe the policy, and a high-level linear Gaussian model with a learned cause Z = tau_1(X) is fit to approximately match interventional push-forward distributions. The paper contributes two theoretical results: Proposition 4.1 (uniqueness of exact constructive transformations up to affine rescaling under an additive-noise SCM with a non-vanishing Fourier density) and Proposition 4.2 (existence of an exact transformation for a special family of additive SCMs), plus an interpretable Gaussian-kernel function class. Experiments on synthetic SCMs, a Pendulum control task, and a simulated robot table tennis task are presented as demonstrations that the method can identify behavioral biases and failure modes.
Significance. If the theoretical guarantees held in the RL settings, this would be a valuable step toward interpretable, causally grounded explanations of nonlinear RL policies. The synthetic experiments provide a machine-checkable-style validation within the model family of Proposition 4.2, and the Pendulum and table tennis findings are plausible and potentially useful qualitative demonstrations. The paper is also commendably explicit about limitations in the Discussion and in Appendix D, including Conjecture D.2 on the failure of identifiability under finite-support intervention priors. The main gap is that the uniqueness theorem is not in force for the RL case studies, and the abstract's claim that uniqueness 'ensuring learned explanations reflect meaningful causal patterns' overstates what is proven for the empirical demonstrations.
major comments (4)
- [Sec. 5.2 and Sec. 3 vs. Prop. 4.1] The uniqueness guarantee (Prop. 4.1) does not apply to the RL experiments as presented. The theorem assumes the low-level model is the additive-noise SCM of Eq. (4.2) with U_pi(1) independent of U_pi(0) and P1 having a non-vanishing Fourier transform, and it requires exact consistency for all shift interventions i in R^{#pi(1)}. In the Pendulum task the environment dynamics are deterministic (Gymnasium's Pendulum-v1 has no state noise), so the exogenous density is degenerate/supported on a subspace and its Fourier transform cannot be non-vanishing on all of R^{#pi(1)}; the deconvolution step in App. B.1 that divides by F[p1] breaks down. Moreover, the experiments intervene only on the action variables delta A_t, so the implemented interventions lie in a proper subspace of R^{#pi(1)}; equality on that subspace only identifies the projection of delta tau_1, not the full function. The abstract's statement that uniqueness 'ensuring learned explanations reflect meaningful causal patterns' therefore does not follow for the RL case studies; these should be described as qualitative demonstrations, with the uniqueness claim restricted to the synthetic model class.
- [App. B.1, Prop. 4.1] The proof of Proposition 4.1 assumes more than the proposition states. It uses the map g = (id - f1)^{-1} and requires that g is well-defined and invertible on the relevant domain, and that delta tau_1 o g is regular enough for its Fourier transform to exist (e.g., integrable or tempered). None of these conditions appear in the statement of Prop. 4.1, which simply says 'additive noise SCM of the form Eq. (4.2)'. In fact, Eq. (4.2) is self-referential for X_pi(1) (the variable appears on both sides), so unique solvability of the structural equations is a nontrivial assumption. The proposition and proof should be aligned by adding explicit invertibility and regularity hypotheses, or by proving these conditions follow from the stated assumptions.
- [Sec. 4.2 and App. D] The uniqueness result concerns exact interventional consistency for all i, whereas the method actually minimizes Lcons plus a normality regularizer using a finite number of sampled interventions. Proposition D.1 extends identifiability to the case where Lcons vanishes exactly and the intervention prior has strictly positive density over the whole space, but neither condition is satisfied in practice. Conjecture D.2 in the same appendix explicitly states that identifiability can fail when the intervention prior vanishes on an open set, and the ablation in App. F.1 shows that the normality regularizer is needed to recover ground truth in high-noise regimes. Consequently, the paper does not establish uniqueness of the learned nTCR solution in the experiments; it establishes uniqueness only for an idealized exact-transformation setting. The authors should either prove a finite-intervention identifiability statement for the actual objective or clearly state in the main text that the uniqueness guarantee does not apply to the optimized approximate solution.
- [Prop. 4.2 and App. B.2] The existence result is narrower than the narrative in the abstract suggests. It covers only low-level models with the special form f0(X_pi(0), X_pi(1)) = h0(X_pi(0)) + B(X_pi(1) - f1(X_pi(1))) and a target Y of the exact form a^T (X_pi(0) - h0(X_pi(0))). Within that family the result is correct and the synthetic validation is appropriate, but it does not imply that exact nonlinear reductions exist for general RL environments. The wording of Proposition 4.1 also says uniqueness of a solution 'if there exists' one; combining the two propositions as 'there exists a unique solution' in the abstract is misleading and should be corrected.
minor comments (5)
- [Abstract] The phrase 'there exists a unique solution that achieves exact interventional consistency' conflates the conditional uniqueness statement (Prop. 4.1) with the restricted existence statement (Prop. 4.2). Please rephrase to distinguish these two results.
- [Eq. (4.2)] Equation (4.2) uses X_pi(1) on both sides of its own structural equation; this is notationally confusing and should be replaced by an explicit fixed-point formulation or a clearer definition of f1.
- [Sec. 4.1, Eq. (4.1)] The expectation in Eq. (4.1) is written as E_{i ~ P(i)}; the same symbol P(i) denotes an interventional distribution elsewhere in the paper. Use a distinct notation for the intervention prior, e.g., P_I, to avoid ambiguity.
- [Fig. 3 caption] The caption says the 'middle plot on the bottom row' shows mean rewards, but the figure has multiple bottom-row panels; please refer to the plots by their explicit labels or positions.
- [App. E.2] The text says interventions are applied for the first 75 steps, while the time axes in Fig. 3 appear to range from 0 to 50. Please clarify whether the plots are downsampled or whether the intervention window differs from 75 steps.
Circularity Check
No significant circularity: the uniqueness and existence proofs are self-contained, and the self-citations and synthetic validation do not reduce the central claims to their inputs.
full rationale
The paper's central derivation, Prop. 4.1, is a self-contained deconvolution argument: assuming an exact constructive transformation exists, equality of target expectations under all shift interventions yields p1 * δτ1∘g = 0, and the non-vanishing Fourier transform of P1 forces δτ1 = 0. This proof does not invoke any fitted value or the authors' prior TCR paper; the framework definitions in Def. 2.3 are reused from [8], but the uniqueness and existence results are proved directly from the stated assumptions. Prop. 4.2 is a constructive existence proof: the maps in Eq. (4.3) are derived and verified, not assumed, and the model class is deliberately chosen so that a linear reduction is exact. The synthetic experiments generate data inside Prop. 4.2's model class and then check that the optimizer recovers the closed-form solution; this is an implementation sanity check, not a circular inference about unseen data. The self-citations to [8] are for the linear TCR baseline and the general TCR formulation; they are not load-bearing for the new nonlinear guarantees. The paper itself flags the scope of its guarantees: Sec. 7 states that the results are limited to the model classes in Props. 4.1 and 4.2, and App. D's Conjecture D.2 acknowledges that identifiability can fail when the intervention prior does not have full support. The gap between Prop. 4.1's 'all i ∈ R^{#π(1)}' requirement and the RL experiments' action-only interventions, as well as the deterministic dynamics in Pendulum, are threats to transferring the uniqueness guarantee to the empirical demonstrations, but these are correctness and applicability concerns, not circularity. No step reduces the derivation to its own inputs.
Assumptions & free parameters
free parameters (4)
- normality regularization weight eta_norm =
1.0 (synthetic), 10 (pendulum)
- kernel smoothing constant c =
8 (pendulum), 1 (table tennis)
- intervention strength sigma =
Selected per task via Fig A threshold
- high-level model parameters alpha and bias =
e.g., alpha=0.0013, bias=0.5964 for nTCR table tennis
assumptions (6)
- domain assumption Low-level model is an additive-noise SCM with Xπ(1) = f1(Xπ(1)) + Uπ(1) + iπ(1) and Xπ(0) = f0(Xπ(0), Xπ(1)) + Uπ(0), with Uπ(1) independent of Uπ(0) and P1 having a density with non-vanishing Fourier transform.
- domain assumption High-level model is a linear additive Gaussian noise model with non-zero causal coefficient alpha.
- domain assumption Action perturbations delta_At can be treated as additive shift interventions on the structural equations of action variables.
- ad hoc to paper For the existence result, f0 has the special form h0(Xπ(0)) + B(Xπ(1) - f1(Xπ(1))) and Y = a^T (Xπ(0) - h0(Xπ(0))).
- standard math Fourier transform and convolution theorems are used to derive uniqueness.
- domain assumption The RL environment can be represented as a low-level SCM with exogenous noise, i.e., initial state and any stochasticity are captured by U.
invented entities (1)
-
High-level cause variable Z = tau1(Xπ(1))
Cite this review
Pith. "Pith review of Learning Nonlinear Causal Reductions to Explain Reinforcement Learning Policies." pith.science (2026). https://pith.science/paper/IUCZFJFH
@misc{pith2026250714901,
author = {Pith},
title = {Pith review of: Learning Nonlinear Causal Reductions to Explain Reinforcement Learning Policies},
year = {2026},
howpublished = {\url{https://pith.science/paper/IUCZFJFH}},
note = {Machine review of arXiv:2507.14901}
}
read the original abstract
Why do reinforcement learning (RL) policies fail or succeed? This is a challenging question due to the complex, high-dimensional nature of agent-environment interactions. In this work, we take a causal perspective on explaining the behavior of RL policies by viewing the states, actions, and rewards as variables in a low-level causal model. We introduce random perturbations to policy actions during execution and observe their effects on the cumulative reward, learning a simplified high-level causal model that explains these relationships. To this end, we develop a nonlinear Causal Model Reduction framework that ensures approximate interventional consistency, meaning the simplified high-level model responds to interventions in a similar way as the original complex system. We prove that for a class of nonlinear causal models, there exists a unique solution that achieves exact interventional consistency, ensuring learned explanations reflect meaningful causal patterns. Experiments on both synthetic causal models and practical RL tasks-including pendulum control and robot table tennis-demonstrate that our approach can uncover important behavioral patterns, biases, and failure modes in trained RL policies.
Figures
Figures from the paper (1 more)
Reference graph
Works this paper leans on
-
[1]
Mastering the game of Go with deep neural networks and tree search
David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of Go with deep neural networks and tree search. Nature, 2016. [Cited on page 1.]
work page 2016
-
[2]
Playing atari with deep reinforcement learning
V olodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv:1312.5602, 2013. [Cited on page 1.]
arXiv 2013
-
[3]
Reinforcement learning in robotics: A survey
Jens Kober, J Andrew Bagnell, and Jan Peters. Reinforcement learning in robotics: A survey. The International Journal of Robotics Research, 2013. [Cited on page 1.]
work page 2013
-
[4]
Resource management with deep reinforcement learning
Hongzi Mao, Mohammad Alizadeh, Ishai Menache, and Srikanth Kandula. Resource management with deep reinforcement learning. In Proceedings of the 15th ACM workshop on hot topics in networks, 2016. [Cited on page 1.]
work page 2016
-
[5]
End to end learning for self-driving cars
Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, et al. End to end learning for self-driving cars. arXiv:1604.07316, 2016. [Cited on page 1.]
arXiv 2016
-
[6]
Reinforcement learning based recommender systems: A survey
M Mehdi Afsar, Trafford Crump, and Behrouz Far. Reinforcement learning based recommender systems: A survey. ACM Computing Surveys, 2022. [Cited on page 1.]
work page 2022
-
[7]
A review on reinforcement learning: Introduction and applications in industrial process control
Rui Nian, Jinfeng Liu, and Biao Huang. A review on reinforcement learning: Introduction and applications in industrial process control. Computers & Chemical Engineering, 2020. [Cited on page 1.]
work page 2020
-
[8]
Targeted Reduction of Causal Models
Armin Keki´c, Bernhard Schölkopf, and Michel Besserve. Targeted Reduction of Causal Models. UAI,
Show all 43 references
-
[9]
Causality
Judea Pearl. Causality. Cambridge university press, second edition, 2009. [Cited on page 3.]
2009
-
[10]
Elements of Causal Inference – Foundations and Learning Algorithms
Jonas Peters, Dominik Janzing, and Bernhard Schölkopf. Elements of Causal Inference – Foundations and Learning Algorithms. MIT Press, 2017. [Cited on page 3.]
2017
-
[11]
Abstracting Causal Models
Sander Beckers and Joseph Y Halpern. Abstracting Causal Models. AAAI, 2019. [Cited on pages 3 and 9.]
2019
-
[12]
Approximate Causal Abstractions
Sander Beckers, Frederick Eberhardt, and Joseph Y Halpern. Approximate Causal Abstractions. UAI, 2020
2020
-
[13]
Causal Abstraction with Soft Interventions
Riccardo Massidda, Atticus Geiger, Thomas Icard, and Davide Bacciu. Causal Abstraction with Soft Interventions. CLeaR, 2023
2023
-
[14]
Compositional abstraction error and a category of causal models
Eigil F Rischel and Sebastian Weichwald. Compositional abstraction error and a category of causal models. UAI, 2021. [Cited on page 9.]
2021
-
[15]
Causal abstraction for faithful model interpretation
Atticus Geiger, Chris Potts, and Thomas Icard. Causal abstraction for faithful model interpretation. arXiv:2301.04709, 2023. [Cited on page 9.]
2023 arXiv
-
[16]
Visual causal feature learning
Krzysztof Chalupka, Pietro Perona, and Frederick Eberhardt. Visual causal feature learning. UAI, 2015. [Cited on pages 3 and 9.]
2015
-
[17]
Causal consistency of structural equation models
Paul K Rubenstein, Sebastian Weichwald, Stephan Bongers, Joris M Mooij, Dominik Janzing, Moritz Grosse-Wentrup, and Bernhard Schölkopf. Causal consistency of structural equation models. UAI, 2017. [Cited on pages 3, 4, and 9.] 10
2017
-
[18]
Homomor- phism Autoencoder–Learning Group Structured Representations from Observed Transitions
Hamza Keurti, Hsiao-Ru Pan, Michel Besserve, Benjamin F Grewe, and Bernhard Schölkopf. Homomor- phism Autoencoder–Learning Group Structured Representations from Observed Transitions. ICML, 2023. [Cited on page 3.]
2023
-
[19]
Gymnasium: A Standard Interface for Reinforcement Learning Environments
Mark Towers, Ariel Kwiatkowski, Jordan Terry, John U Balis, Gianluca De Cola, Tristan Deleu, Manuel Goulão, Andreas Kallinteris, Markus Krimmel, Arjun KG, et al. Gymnasium: A Standard Interface for Reinforcement Learning Environments. arXiv:2407.17032, 2024. [Cited on pages 7 and 20.]
2024 arXiv
-
[20]
Learning to play table tennis from scratch using muscular robots
Dieter Büchler, Simon Guist, Roberto Calandra, Vincent Berenz, Bernhard Schölkopf, and Jan Peters. Learning to play table tennis from scratch using muscular robots. IEEE Transactions on Robotics, 2022. [Cited on pages 7 and 21.]
2022
-
[21]
Safe & accurate at speed with tendons: A robot arm for exploring dynamic motion
Simon Guist, Jan Schneider, Hao Ma, Le Chen, Vincent Berenz, Julian Martus, Heiko Ott, Felix Grüninger, Michael Muehlebach, Jonathan Fiene, Bernhard Schölkopf, and Dieter Büchler. Safe & accurate at speed with tendons: A robot arm for exploring dynamic motion. In Robotics: Sci...
2024
-
[22]
Explainable reinforcement learning: A survey and comparative review
Stephanie Milani, Nicholay Topin, Manuela Veloso, and Fei Fang. Explainable reinforcement learning: A survey and comparative review. ACM Computing Surveys, 2024. [Cited on page 9.]
2024
-
[23]
Visualizing and understanding Atari agents
Samuel Greydanus, Anurag Koul, Jonathan Dodge, and Alan Fern. Visualizing and understanding Atari agents. ICML, 2018. [Cited on pages 9 and 10.]
2018
-
[24]
Transparency and explanation in deep reinforcement learning neural networks
Rahul Iyer, Yuezhang Li, Huao Li, Michael Lewis, Ramitha Sundar, and Katia Sycara. Transparency and explanation in deep reinforcement learning neural networks. In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, 2018
2018
-
[25]
Towards interpretable reinforcement learning using attention augmented agents
Alexander Mott, Daniel Zoran, Mike Chrzanowski, Daan Wierstra, and Danilo Jimenez Rezende. Towards interpretable reinforcement learning using attention augmented agents. NeurIPS, 2019. [Cited on page 9.]
2019
-
[26]
Explainable robotic systems: Under- standing goal-driven actions in a reinforcement learning scenario
Francisco Cruz, Richard Dazeley, Peter Vamplew, and Ithan Moreira. Explainable robotic systems: Under- standing goal-driven actions in a reinforcement learning scenario. Neural Computing and Applications,
-
[27]
Explaining reinforcement learning to mere mortals: An empirical study
Andrew Anderson, Jonathan Dodge, Amrita Sadarangani, Zoe Juozapaitis, Evan Newman, Jed Irvine, Souti Chattopadhyay, Alan Fern, and Margaret Burnett. Explaining reinforcement learning to mere mortals: An empirical study. IJCAI, 2019
2019
-
[28]
Learning "what-if" explanations for sequential decision-making
Ioana Bica, Daniel Jarrett, Alihan Hüyük, and Mihaela van der Schaar. Learning "what-if" explanations for sequential decision-making. ICLR, 2021. [Cited on page 9.]
2021
-
[29]
Graying the black box: Understanding DQNs
Tom Zahavy, Nir Ben-Zrihem, and Shie Mannor. Graying the black box: Understanding DQNs. ICML,
-
[30]
Generation of policy-level explanations for reinforcement learning
Nicholay Topin and Manuela Veloso. Generation of policy-level explanations for reinforcement learning. AAAI, 2019
2019
-
[31]
TLdR: Policy summarization for factored SSP problems using temporal abstractions
Sarath Sreedharan, Siddharth Srivastava, and Subbarao Kambhampati. TLdR: Policy summarization for factored SSP problems using temporal abstractions. ICAPS, 2020. [Cited on page 9.]
2020
-
[32]
Explainable reinforcement learning through a causal lens
Prashan Madumal, Tim Miller, Liz Sonenberg, and Frank Vetere. Explainable reinforcement learning through a causal lens. AAAI, 2020. [Cited on page 9.]
2020
-
[33]
Causal abstractions of neural networks
Atticus Geiger, Hanson Lu, Thomas Icard, and Christopher Potts. Causal abstractions of neural networks. NeurIPS, 2021. [Cited on page 9.]
2021
-
[34]
Segment anything
Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. ICCV, 2023. [Cited on page 10.]
2023
-
[35]
Grad-CAM: Visual explanations from deep networks via gradient-based localization
Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual explanations from deep networks via gradient-based localization. ICCV,
-
[36]
Foundations of structural causal models with cycles and latent variables
Stephan Bongers, Patrick Forré, Jonas Peters, and Joris M Mooij. Foundations of structural causal models with cycles and latent variables. The Annals of Statistics, 2021. [Cited on pages 13 and 16.]
2021
-
[37]
Dependence, correlation and gaussianity in independent component analysis
Jean-François Cardoso. Dependence, correlation and gaussianity in independent component analysis. JMLR, 2003. [Cited on page 14.] 11
2003
-
[38]
Stable-Baselines 3: Reliable reinforcement learning implementations
Antonin Raffin, Ashley Hill, Adam Gleave, Anssi Kanervisto, Maximilian Ernestus, and Noah Dormann. Stable-Baselines 3: Reliable reinforcement learning implementations. JMLR, 2021. [Cited on pages 21 and 22.]
2021
-
[39]
Proximal policy optimization algorithms
John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv:1707.06347, 2017. [Cited on pages 21 and 22.] 12 Appendix Contents A Additional Background 13 A.1 Cyclic Structural Causal Models . . . . . . . . . . ....
2017 arXiv
-
[41]
The high-level model has (n+1) endogenous variables {Y, Z1,
Target-oriented Structure: We designate a scalar target variable Y = τ0(X) that quantifies a phenomenon of interest. The high-level model has (n+1) endogenous variables {Y, Z1, . . . , Zn}, where Z1, . . . , Zn are the learned high-level causes of Y
-
[42]
The exogenous variables {W0, W1,
Parameterized High-level Models : The high-level SCM is constrained to a class of linear additive Gaussian noise models {Hγ }γ∈Γ with parameters γ to be learned. The exogenous variables {W0, W1, . . . , Wn} have a factorized Gaussian distribution PW = Q PWk
-
[43]
(id − f1)−1
Constructive Transformations: The linear maps τ and ω are constructive, meaning each dimension depends only on a designated subset of low-level variables: τ = (τ0, τ1, . . . , τn) with τk : x 7→ ¯τk(xπ(k)) (A.1) ω = (ω0, ω1, . . . , ωn) with ωk : i 7→ ¯ωk(iπ(k)) . (A.2) Here, ...
-
[2024]
[Cited on pages 2, 3, 4, 5, 9, 14, 15, 17, 18, 24, and 25.]
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.