Pith. sign in

REVIEW 4 major objections 5 minor 56 references

CASTLE: Regularization via Auxiliary Causal Graph Discovery

T0 review · 4 major / 5 minor · reviewed 2026-08-27 · deepseek-v4-flash

Pith's one-line read CASTLE regularizes neural networks by jointly learning the causal DAG among features and target, claiming better out-of-sample generalization than standard regularizers.

desk verdict CASTLE is a promising and pragmatic regularizer for causal tabular data, but the core DAG mechanism is mis-specified in the text and the 'never degrades' claim is overreach. read the letter →

arxiv 2009.13180 v1 pith:MN2FMDTH submitted 2020-09-28 cs.LG stat.ML

classification cs.LGstat.ML
keywords causalregularizationDAGdiscoveryneuralnetworkgeneralizationboundsupervisedautoencodersnonparametricstructuralequationmodelacyclicityconstraintout-of-sampleprediction
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

Causal regularization, not just capacity control, is the route to better out-of-sample predictions on tabular data: the paper introduces CASTLE, which trains a neural network while simultaneously learning the directed acyclic graph (DAG) among the features and the target. The same objective that predicts the target also reconstructs only those variables that have causal neighbors, and a continuous acyclicity penalty keeps the learned adjacency matrix acyclic. The paper argues this gives two advantages over existing regularizers: it selects causal parents rather than merely correlated features, and it avoids wasting capacity on reconstructing noise variables that standard supervised autoencoders reconstruct. It backs the claim with a PAC-Bayes generalization bound and experiments on synthetic DAGs and real datasets where CASTLE consistently ranks first.

What carries the argument

The load-bearing object is the adjacency matrix $M$ embedded in the input-layer weights of a feed-forward network with $d+1$ sub-networks: $[M]_{k,j}$ is the $\ell^2$-norm of the $k$-th row of sub-network $j$'s input weight matrix, so it records the strength of edge $k\to j$. The mechanism that carries the argument is the continuous acyclicity penalty $R_{\Theta_1}=(\mathrm{Tr}(e^{M\odot M})-d-1)^2$, which is zero exactly when the learned graph is a DAG, combined with the reconstruction loss $L_N(f_\Theta)=\frac{1}{N}\|\tilde{X}-f_\Theta(\tilde{X})\|_F^2$ that reconstructs each variable only from the others. The PAC-Bayes bound in Theorem 1 reinterprets this combination as a prior on the input weights and controls the expected reconstruction loss by the empirical loss plus $(R_{\Theta_1}+C_1(V_{\Theta_1}+V_{\Theta_2})+\log(8/\delta))/N$, showing the acyclicity term acts as a capacity term with its largest effect at small sample sizes.

What would settle it

Simulate data from a known DAG in which the target is an exact invertible function of one of its own descendants, so an anti-causal predictor beats the true parents; if CASTLE still improves out-of-sample error without assigning weight to the true parents, the causal-DAG mechanism is not what produces the gain. Conversely, on data with no causal structure at all, independent features unrelated to the target, CASTLE should not beat early stopping; if it does, the benefit is not tied to the DAG assumption.

Watch

Extended reading notes

Core claim

The central claim is that a supervised network generalizes better when the training objective itself discovers the causal DAG among $[Y,X]$ rather than using a causality-agnostic penalty. CASTLE's objective is $\min_\Theta \frac{1}{N}\|Y-[f_\Theta(\tilde{X})]_{:,1}\|^2 + \lambda R_{DAG}(\tilde{X},f_\Theta)$, where the regularizer is the sum of a neighbor-only reconstruction loss, an $\ell^1$ sparsity term, and the acyclicity penalty $R_{\Theta_1}=(\mathrm{Tr}(e^{M\odot M})-d-1)^2$, with $M$ encoding the $\ell^2$ norms of the input-layer weight rows. The paper asserts that minimizing this joint objective achieves both causal prediction and target selection, and reports lower test error than early stopping, L1, L2, dropout, batch normalization, input noise, MixUp, and supervised autoencoders across its synthetic and real benchmark datasets.

Load-bearing premise

The regularizer only helps when the observed features and target really are generated by an underlying causal DAG and the continuous acyclicity penalty recovers that DAG from finite data; where that premise fails, as the paper itself concedes for image data, CASTLE has no grounding and can degrade predictions.

Editorial extensions

If this is right

  • On tabular data with an underlying causal structure, CASTLE should give lower out-of-sample error than capacity-based regularizers and supervised autoencoders at matched architectures.
  • Reconstruction-based regularization should reconstruct only variables with causal neighbors, because reconstructing noise variables, as standard autoencoders do, can actively hurt generalization.
  • The PAC-Bayes bound predicts the benefit of the acyclicity term shrinks as sample size grows, so CASTLE's advantage should be largest in small-$N$ regimes.
  • Causal parents are the preferred predictors when they exist; when they do not, the method falls back to children of the target rather than unrelated correlated features.
  • CASTLE can be combined with conventional capacity control ($V_{\Theta_2}$) since the bound includes both terms, potentially giving further gains.

Reading between the lines

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

  • The paper's generalization bound is stated for the reconstruction loss of the whole vector $\tilde{X}$, not directly for the prediction loss of $Y$; interpreting the bound as a guarantee for the supervised task requires an extra step the paper leaves implicit.
  • A natural stress test the authors do not run: vary the proportion of noise (neighborless) variables while keeping the true parents of $Y$ weak; CASTLE should separate from SAE exactly where target selection matters most.
  • The computational cost of training the joint network grows sharply with the number of features, so a practical successor would need a theory for the feature-subsampling trick the paper mentions but does not analyze.
  • Because the method assumes no hidden confounders, an immediate extension would be to incorporate latent-variable causal discovery; without it, CASTLE may learn a DAG over proxies rather than the true causal graph.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes CASTLE, a regularization method that jointly learns a causal DAG over the input features and target variable while training a neural network for supervised prediction. The DAG is embedded in the input-layer weights of a multi-output network, with a continuous acyclicity penalty and a sparsity penalty used as the regularizer, and the auxiliary task is reconstruction of only the variables that have causal neighbors. The authors provide a PAC-Bayes generalization bound for the reconstruction loss, an ablation study, and large-scale experiments on synthetic and real tabular datasets comparing CASTLE with capacity-based and reconstruction-based regularizers. The central claim is that this causal inductive bias improves out-of-sample prediction and target selection relative to existing regularizers.

Significance. The core idea—using causal structure learning as an inductive bias for regularization and for selecting which variables to reconstruct—is original and potentially valuable. If the formal mechanism is made precise, the paper could make a useful contribution to the intersection of causality and deep learning regularization. The authors provide code, extensive experiments across many datasets, and an explicit attempt at a theoretical bound. However, the current manuscript has a load-bearing mathematical inconsistency in the definition of the DAG regularizer, and the theoretical bound addresses reconstruction loss rather than the supervised prediction loss that is the paper's headline. These issues condition the significance on corrections rather than establishing the claims as written.

major comments (4)
  1. [Section 3.3, Eq. (3), Eq. (7)] The dimensions and indexing of the input weight matrices W_k^1 are internally inconsistent. The text declares W_k^1 to be an h x (d+1) matrix, but the definition f_k(\tilde X) = φ(...φ(\tilde X W_k^1)...) requires W_k^1 to have d+1 rows so that the product with the N x (d+1) input matrix is defined. With the declared h x (d+1) shape, the row dimension of W_j^1 runs over hidden units, so [M]_{k,j} = ||row_k(W_j^1)||_2 in Eq. (7) does not define a (d+1)-variable adjacency matrix; it mixes hidden-unit indices with variable indices. The masking instruction "set the k-th column of W_k^1 to zero" is also incompatible with the only shape that makes Eq. (3) well-defined, where the k-th row must be zeroed. As written, the manuscript does not formally define the causal DAG adjacency matrix that is claimed to be embedded in the network, so the central mechanism of CASTLE is not checkable from the text alone. This should be corrected and the derivation of Eq. (7) restated consistently; the released code may resolve the intended convention, but the paper must be self-contained.
  2. [Section 3.4, Theorem 1, Eq. (8)] The theoretical result bounds the expected reconstruction loss L_P(f_Θ) = E||f_Θ(\tilde X) - \tilde X||^2, not the supervised prediction loss for Y. The abstract and introduction claim that CASTLE improves out-of-sample prediction of Y, but the bound contains no term for the prediction error of the first output [f_Θ]_{:,1}. The paragraph following Eq. (8) asserts that the reconstruction-loss result "explains the benefit" for predicting Y, but no Lipschitz or other transfer argument is given to connect reconstruction accuracy of the joint vector to prediction accuracy of its first component. If the theorem is meant to support the generalization of the prediction task, an additional argument or a modified bound is needed; otherwise the claim should be narrowed explicitly.
  3. [Appendix A, Eqs. (14)-(15)] The KL-divergence bound appears to have an algebraic inconsistency with the final theorem statement. Eq. (13) gives L_P ≤ 4L_N + (2/N)[2 KL + log(8/δ)] + C_2, which has coefficient 4/N on the KL term. Substituting the bound KL ≤ (1/2)R_θ1 + (1/(2σ^2))(V_Θ1+V_Θ2) + ... would yield a coefficient of 2/N for R_θ1 in the final display, not 1/N as printed in Eq. (15) and Theorem 1. Please recheck the constants and ensure that the displayed bound follows from the preceding display; the current proof does not verify the theorem as stated.
  4. [Section 5 and Tables 3, 5, 6] The conclusion states that "in the worst case, never degrades performance," but the empirical evidence is based on means and standard deviations without any paired significance tests across folds or datasets. In several cases the CASTLE improvement is within one standard deviation of the baseline (e.g., BC regression 0.318±0.036 vs. Baseline 0.359±0.061; LV classification 0.595±0.032 vs. Baseline 0.559±0.061). The "never degrades" claim is stronger than what is demonstrated; either provide appropriate statistical tests (e.g., paired tests over folds) or soften the claim to the observed datasets and conditions.
minor comments (5)
  1. [Section 3.3, "Target Selection"] The statement that CASTLE "reconstructs only the features in the causal DAG that have a causal neighbor" is not a formal property of the objective: the reconstruction loss L_N(f_Θ) attempts to reconstruct every variable except the self-masked one, and the near-zero weights for noise variables arise from the L1 penalty and the optimization, not from an explicit constraint. Consider presenting this as an empirical property rather than a structural guarantee.
  2. [Section 4.1, Dissecting CASTLE] The text says "the best and worst possible rank is one and r, respectively (i.e., the higher the rank the better)"; if rank 1 is best, a higher numeric rank is worse. Please clarify the direction of the ranking or correct the wording, as the figure caption and accompanying discussion are ambiguous.
  3. [Theorem 1 in Section 3.4] The displayed bound in Eq. (8) uses a constant C_3, while the surrounding text refers to C_1 and C_2, and the appendix version of the theorem uses C_2. Unify the notation to avoid confusion.
  4. [Section 4, Experimental setup] The preprocessing description says all continuous variables are standardized, but the classification datasets surely contain categorical variables; please specify the encoding (e.g., one-hot) and whether those were standardized, since the DAG and the weight-matrix mechanism depend on the input representation.
  5. [Appendix C.1, Sensitivity analysis] The sensitivity analysis for λ is performed on synthetic DAGs only, yet λ = 1 is then used for all real datasets; since the scale of the reconstruction loss and the acyclicity penalty may vary across datasets even after standardization, a brief justification or a real-data sensitivity check would strengthen the hyperparameter choice.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DAG regularizer is an external, independently grounded inductive bias and the generalization bound is a standard complexity-style bound, not a fitted quantity.

full rationale

CASTLE is a regularization method that adds a DAG-structure penalty to a supervised objective. The central elements are (a) the continuous acyclicity penalty from the external works [37, 38], (b) a reconstruction loss over all variables, and (c) a PAC-Bayes generalization bound in which the DAG penalty appears as a complexity term. None of these reduces to the paper's own empirical claim by construction. The acyclicity penalty R_Theta1 is not defined in terms of the target prediction or generalization performance; it is a function of the learned weight matrices and is imported from prior work by Zheng et al., not from the present authors. The objective in Eq. (5) genuinely combines a prediction loss with a regularizer, and the experiments compare against independent benchmarks, so there is no fitted parameter that is then renamed as a prediction. The PAC-Bayes bound in Theorem 1 is valid for all parameters simultaneously and treats R_Theta1 as part of the complexity term; minimizing it lowers an upper bound, but this is standard regularization theory rather than a circular equation. The 'target selection' property, i.e., reconstructing only variables with neighbors in the learned DAG, follows from the sparsity and acyclicity penalties rather than being assumed as the conclusion. The citations to the authors' own earlier work are not load-bearing for the main derivation. There is a separate formal clarity concern about the stated dimensions of W_k^1 and the indexing in Eq. (7), but that is an internal consistency issue, not a circular reduction of the claimed result to its inputs. Overall, no circular step meeting the required standard was found.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim depends on two tuned scalar hyperparameters and on the validity of the DAG/NPSEM assumption. No new physical or conceptual entities are introduced; the adjacency matrix is a parameterization of an existing idea.

free parameters (2)
  • lambda (DAG loss weight) = 1
    Chosen via sensitivity analysis in Appendix C.1 over {0.01, 0.1, 1, 10, 100}; fixed for all experiments.
  • beta (sparsity penalty) = Not reported per dataset; grid searched over {0.001, 0.01, 0.1, 1}
    Controls the l1 penalty on input adjacency matrices in Eq. (6); authors do not state which value was used for each result.
assumptions (4)
  • domain assumption The variables [Y,X] satisfy a nonparametric structural equation model with an underlying DAG G (Definition 1).
    The causal DAG is the central inductive bias; without it the acyclicity penalty has no target to discover.
  • standard math A matrix W is a DAG if and only if Tr(e^{W⊙W}) - (d+1) = 0 (Theorem 1 in [37]).
    Borrowed from NOTEARS; the continuous acyclicity constraint is used wholesale in Equations (2) and (6).
  • domain assumption The squared reconstruction loss is sub-Gaussian under the data distribution (Assumption 2).
    Needed for the PAC-Bayes argument; plausible for bounded data but not proven for the datasets used.
  • standard math PAC-Bayes generalization bounds and the derandomization of Neyshabur et al. [45] apply to deterministic ReLU networks.
    Base of Theorem 1; the authors rely on existing theory rather than deriving new concentration inequalities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CASTLE: Regularization via Auxiliary Causal Graph Discovery." pith.science (2026). https://pith.science/paper/MN2FMDTH

@misc{pith2026200913180,
  author       = {Pith},
  title        = {Pith review of: CASTLE: Regularization via Auxiliary Causal Graph Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MN2FMDTH}},
  note         = {Machine review of arXiv:2009.13180}
}
read the original abstract

Regularization improves generalization of supervised models to out-of-sample data. Prior works have shown that prediction in the causal direction (effect from cause) results in lower testing error than the anti-causal direction. However, existing regularization methods are agnostic of causality. We introduce Causal Structure Learning (CASTLE) regularization and propose to regularize a neural network by jointly learning the causal relationships between variables. CASTLE learns the causal directed acyclical graph (DAG) as an adjacency matrix embedded in the neural network's input layers, thereby facilitating the discovery of optimal predictors. Furthermore, CASTLE efficiently reconstructs only the features in the causal DAG that have a causal neighbor, whereas reconstruction-based regularizers suboptimally reconstruct all input features. We provide a theoretical generalization bound for our approach and conduct experiments on a plethora of synthetic and real publicly available datasets demonstrating that CASTLE consistently leads to better out-of-sample predictions as compared to other popular benchmark regularizers.

Figures

Figures reproduced from arXiv: 2009.13180 by the authors.

Figure 1
Figure 1. Example DAG. Target Selection. Consider the variables X5, X6 and X7 which share parents (X2 and X3) with Y in [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Schematic of CASTLE regularization. Our goal is to have the following tasks: (1) a [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Experiments on synthetic data. The y-axis is the average rank (± standard deviation) of each regularizer on the test set over each synthetic DAG. We show the average rank as we increase the number of features or vertex cardinality |G| (left), increase the dataset size normalized by the vertex cardinality |G| (center), and as we increase the number of noise (neighborless) variables (right). Benchmarks. We benchmark C… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Sensitivity analysis on λ. Before we present further results, we first provide a sensitivity analysis on λ from (5). We use our synthetic DGP to synthesize a random DAG with between 10 and 150 nodes. We generated 2000 test samples and a training set with between 1000 a…
Figure 5
Figure 5. Figure 5: CASTLE scalability analysis We perform an analysis of the scalability of CASTLE. Using our synthetic DAG and dataset generator, we synthesized datasets of 1000 samples. We used the same experimental setup used for the synthetic experiments. We present the computational…
Figure 6
Figure 6. Figure 6: Comparison of CASTLE against benchmark regularization methods in terms or average [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: Weight values on synthetic data when true causal structure is known. Our method favors [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: Weight values on synthetic data when true causal structure is known. This simulation [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 54 canonical work pages

  1. [1]

    Yaeger, Richard F

    Larry S. Yaeger, Richard F. Lyon, and Brandyn J. Webb. Effective training of a neural network character classifier for word recognition. In M. C. Mozer, M. I. Jordan, and T. Petsche, editors, Advances in Neural Information Processing Systems 9, pages 807–816. MIT Press, 1997

  2. [2]

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Proceedings of the 25th International Conference on Neural Information Processing Systems - Volume 1, NIPS’12, page 1097–1105, Red Hook, NY , USA,

  3. [3]

    Improving neural networks by preventing co-adaptation of feature detectors

    Geoffrey E. Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhut- dinov. Improving neural networks by preventing co-adaptation of feature detectors. ArXiv, abs/1207.0580, 2012

  4. [4]

    Dropout training as adaptive regularization

    Stefan Wager, Sida Wang, and Percy Liang. Dropout training as adaptive regularization. Advances in Neural Information Processing Systems, 07 2013

  5. [5]

    Dropout: A simple way to prevent neural networks from overfitting

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(56):1929–1958, 2014

  6. [6]

    Adversarial regularizers in inverse problems

    Sebastian Lunz, Ozan Öktem, and Carola-Bibiane Schönlieb. Adversarial regularizers in inverse problems. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and 9 R. Garnett, editors, Advances in Neural Information Processing Systems 31, pages 8507–8516. Curran Associates, Inc., 2018

  7. [7]

    Deep label distribution learning with label ambiguity

    Bin-Bin Gao, Chao Xing, Chen-Wei Xie, Jianxin Wu, and Xin Geng. Deep label distribution learning with label ambiguity. IEEE Transactions on Image Processing, 26:2825–2838, 04 2017

  8. [8]

    Greedy layer-wise training of deep networks

    Yoshua Bengio, Pascal Lamblin, Dan Popovici, and Hugo Larochelle. Greedy layer-wise training of deep networks. In Proceedings of the 19th International Conference on Neural Information Processing Systems, NIPS’06, page 153–160, Cambridge, MA, USA, 2006. MIT Press

Show all 56 references
  1. [9]

    Semi-supervised learning of compact document representations with deep networks

    Marc’Aurelio Ranzato and Martin Szummer. Semi-supervised learning of compact document representations with deep networks. In Proceedings of the 25th International Conference on Machine Learning, pages 792–799, 01 2008

  2. [10]

    Layer-wise coordination between encoder and decoder for neural machine translation

    Tianyu He, Xu Tan, Yingce Xia, Di He, Tao Qin, Zhibo Chen, and Tie-Yan Liu. Layer-wise coordination between encoder and decoder for neural machine translation. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Inf...

  3. [11]

    Extracting and composing robust features with denoising autoencoders

    Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. InProceedings of the 25th international conference on Machine learning, pages 1096–1103. ACM, 2008

  4. [12]

    Supervised autoencoders: Improving generaliza- tion performance with unsupervised regularizers

    Lei Le, Andrew Patterson, and Martha White. Supervised autoencoders: Improving generaliza- tion performance with unsupervised regularizers. In Advances in Neural Information Processing Systems, pages 107–117, 2018

  5. [13]

    Causal discovery with reinforcement learning

    Shengyu Zhu and Zhitang Chen. Causal discovery with reinforcement learning. CoRR, abs/1906.04477, 2019

  6. [14]

    Triad constraints for learning causal structure of latent variables

    Ruichu Cai, Feng Xie, Clark Glymour, Zhifeng Hao, and Kun Zhang. Triad constraints for learning causal structure of latent variables. In Advances in Neural Information Processing Systems 32, pages 12883–12892. Curran Associates, Inc., 2019

  7. [15]

    Estimating individual treatment effect: generalization bounds and algorithms

    Uri Shalit, Fredrik D Johansson, and David Sontag. Estimating individual treatment effect: generalization bounds and algorithms. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3076–3085. JMLR. org, 2017

  8. [16]

    Validating causal inference models via influence functions

    Ahmed Alaa and Mihaela van der Schaar. Validating causal inference models via influence functions. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 ofProceedings of Machine Learning Research...

  9. [17]

    Cxplain: Causal explanations for model interpretation under uncertainty

    Patrick Schwab and Walter Karlen. Cxplain: Causal explanations for model interpretation under uncertainty. In Advances in Neural Information Processing Systems 32, pages 10220–10230. Curran Associates, Inc., 2019

  10. [18]

    Domain adaptation un- der target and conditional shift

    Kun Zhang, Bernhard Schölkopf, Krikamol Muandet, and Zhikun Wang. Domain adaptation un- der target and conditional shift. In Sanjoy Dasgupta and David McAllester, editors,Proceedings of the 30th International Conference on Machine Learning (ICML), volume 28 of Proceedings of M...

  11. [19]

    Causal inference by using invariant prediction: identification and confidence intervals

    Jonas Peters, Peter Bühlmann, and Nicolai Meinshausen. Causal inference by using invariant prediction: identification and confidence intervals. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 78(5):947–1012, 2016

  12. [20]

    Domain adaptation by using causal inference to predict invariant conditional distributions

    Sara Magliacane et al. Domain adaptation by using causal inference to predict invariant conditional distributions. In S. Bengio et al., editors,Advances in Neural Information Processing Systems 31, pages 10846–10856. Curran Associates, Inc., 2018

  13. [21]

    Invariant models for causal transfer learning

    Mateo Rojas-Carulla, Bernhard Schölkopf, Richard Turner, and Jonas Peters. Invariant models for causal transfer learning. Journal of Machine Learning Research, 19(36):1–34, 2018. 10

  14. [22]

    On causal and anticausal learning

    Bernhard Schoelkopf, Dominik Janzing, Jonas Peters, Eleni Sgouritsa, Kun Zhang, and Joris Mooij. On causal and anticausal learning. Proceedings of the 29th International Conference on Machine Learning, ICML 2012, 2, 06 2012

  15. [23]

    Stewart, and Jimeng Sun

    Mohammad Taha Bahadori, Krzysztof Chalupka, Edward Choi, Robert Chen, Walter F. Stewart, and Jimeng Sun. Causal regularization. CoRR, abs/1702.02604, 2017

  16. [24]

    Anchor regression: heterogeneous data meet causality

    Dominik Rothenhausler, Nicolai Meinshausen, Peter Buhlmann, and Jonas Peters. Anchor regression: heterogeneous data meet causality. CoRR, abs/1801.06229, 2018

  17. [25]

    Causal regularization

    Dominik Janzing. Causal regularization. In Advances in Neural Information Processing Systems 32, pages 12704–12714. Curran Associates, Inc., 2019

  18. [26]

    On causal and anticausal learning

    Bernhard Schölkopf, Dominik Janzing, Jonas Peters, Eleni Sgouritsa, Kun Zhang, and Joris Mooij. On causal and anticausal learning. In Proceedings of the 29th International Coference on International Conference on Machine Learning, ICML’12, page 459–466, Madison, WI, USA, 2012....

  19. [27]

    Discovering causal signals in images

    David Lopez-Paz, Robert Nishihara, Soumith Chintala, Bernhard Scholkopf, and Leon Bottou. Discovering causal signals in images. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017

  20. [28]

    Semi-supervised interpolation in an anticausal learning scenario

    Dominik Janzing and Bernhard Schölkopf. Semi-supervised interpolation in an anticausal learning scenario. J. Mach. Learn. Res., 16(1):1923–1948, January 2015

  21. [29]

    Regression shrinkage and selection via the lasso

    Robert Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society: Series B (Methodological), 58(1):267–288, 1996

  22. [30]

    Ridge regression: Biased estimation for nonorthogonal problems

    Arthur Hoerl and Robert Kennard. Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12:55–67, 04 2012

  23. [31]

    Deep Learning

    Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016

  24. [32]

    On the difficulty of training recurrent neural networks

    Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural networks. In ICML, 2012

  25. [33]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift

    Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37 , ICML’15, page 448–456. JMLR.org, 2015

  26. [34]

    Regularizing deep neural networks by noise: Its interpretation and optimization

    Hyeonwoo Noh, Tackgeun You, Jonghwan Mun, and Bohyung Han. Regularizing deep neural networks by noise: Its interpretation and optimization. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, page 5115–5124, Red Hook, NY , USA...

  27. [35]

    mixup: Beyond empirical risk minimization

    Hongyi Zhang, Moustapha Cisse, Yann Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. Proceedings of the 6th International Conference on Learning Representations (ICLR), 2018

  28. [36]

    Estimating causal direction and confounding of two discrete variables

    Krzysztof Chalupka, Frederick Eberhardt, and Pietro Perona. Estimating causal direction and confounding of two discrete variables. arXiv preprint arXiv:1611.01504, 2016

  29. [37]

    Xun Zheng, Bryon Aragam, Pradeep Ravikumar, and Eric P. Xing. DAGs with NO TEARS: Continuous Optimization for Structure Learning. In Advances in Neural Information Processing Systems, 2018

  30. [38]

    Learning sparse nonparametric dags

    Xun Zheng, Chen Dan, Bryon Aragam, Pradeep Ravikumar, and Eric P Xing. Learning sparse nonparametric dags. arXiv preprint arXiv:1909.13189, 2019

  31. [39]

    Gradient- based neural DAG learning

    Sébastien Lachapelle, Philippe Brouillard, Tristan Deleu, and Simon Lacoste-Julien. Gradient- based neural DAG learning. In Proceedings of the 8th International Conference on Learning Representations (ICLR), 2020

  32. [40]

    J. Pearl. Causality. Causality: Models, Reasoning, and Inference. Cambridge Univ. Press, 2009. 11

  33. [41]

    The benefit of multitask representation learning

    Andreas Maurer, Massimiliano Pontil, and Bernardino Romera-Paredes. The benefit of multitask representation learning. The Journal of Machine Learning Research, 17(1):2853–2884, 2016

  34. [42]

    Pac-bayes & margins

    John Langford and John Shawe-Taylor. Pac-bayes & margins. InAdvances in neural information processing systems, pages 439–446, 2003

  35. [43]

    A pac analysis of a bayesian estimator

    John Shawe-Taylor and Robert C Williamson. A pac analysis of a bayesian estimator. In Proceedings of the tenth annual conference on Computational learning theory, pages 2–9, 1997

  36. [44]

    McAllester

    David A. McAllester. Some pac-bayesian theorems. In Machine Learning, pages 230–234. ACM Press, 1998

  37. [45]

    A pac-bayesian approach to spectrally-normalized margin bounds for neural networks

    Behnam Neyshabur, Srinadh Bhojanapalli, and Nathan Srebro. A pac-bayesian approach to spectrally-normalized margin bounds for neural networks. arXiv preprint arXiv:1707.09564, 2017

  38. [46]

    Deterministic pac-bayesian generalization bounds for deep networks via generalizing noise-resilience

    Vaishnavh Nagarajan and J Zico Kolter. Deterministic pac-bayesian generalization bounds for deep networks via generalizing noise-resilience. arXiv preprint arXiv:1905.13344, 2019

  39. [47]

    High-dimensional learning of linear causal networks via inverse covariance estimation

    Po-Ling Loh and Peter Bühlmann. High-dimensional learning of linear causal networks via inverse covariance estimation. The Journal of Machine Learning Research, 15(1):3065–3105, 2014

  40. [48]

    Learning directed acyclic graphs with penalized neighbourhood regression

    Bryon Aragam, Arash A Amini, and Qing Zhou. Learning directed acyclic graphs with penalized neighbourhood regression. arXiv preprint arXiv:1511.08963, 2015

  41. [49]

    The Annals of Statistics, 41(2):536–567, 2013

    Sara Van de Geer, Peter Bühlmann, et al.l0-penalized maximum likelihood for sparse directed acyclic graphs. The Annals of Statistics, 41(2):536–567, 2013

  42. [50]

    UCI machine learning repository, 2020

    Dheeru Dua and Casey Graff. UCI machine learning repository, 2020

  43. [51]

    On large-batch training for deep learning: Generalization gap and sharp minima

    Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016

  44. [52]

    Pac-bayesian theory meets bayesian inference

    Pascal Germain, Francis Bach, Alexandre Lacoste, and Simon Lacoste-Julien. Pac-bayesian theory meets bayesian inference. In Advances in Neural Information Processing Systems, pages 1884–1892, 2016

  45. [53]

    User-friendly tail bounds for sums of random matrices

    Joel A Tropp. User-friendly tail bounds for sums of random matrices. Foundations of computa- tional mathematics, 12(4):389–434, 2012. 12 A Proof of Theorem 1 In this paper, we consider learning a causal DAG as our regularizer. We use a squared loss in our objective function. W...

  46. [54]

    Medical expenditure panel survey (meps), 2020

    Agency for Healthcare Research and Quality. Medical expenditure panel survey (meps), 2020

  47. [55]

    Wong et al

    Chih M. Wong et al. Heart failure in younger patients: the Meta-analysis Global Group in Chronic Heart Failure (MAGGIC). European Heart Journal, 35(39):2714–2721, 06 2014. 20

  48. [2012]

    Curran Associates Inc

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.