acceptodds
Under review as a conference paper at ICLR 2027

Disc-InfoMax: Self-Supervised Learning of Invariant Causal Representations

Abstract

InfoMax self-supervised learning (SSL) has driven significant advances in representation learning, but standard approaches often entangle true causal variables with spurious correlations, leading to catastrophic failures under distribution shifts. While multi-view causal representation learning offers theoretical guarantees for identifying true causal variables, practical implementations in SSL settings are bottlenecked by their reliance on explicit view labels and highly non-convex optimization. To bridge this gap, we address the broader challenge of causal representation learning without explicit view labels, which includes the InfoMax SSL setting. We make two primary contributions. First, we introduce a novel training objective, Disc-InfoMax, which dynamically infers a shared content mask from the empirical differences between paired representations. We prove that under asymptotic convergence, the optimal encoder is theoretically guaranteed to block-identify true underlying causal variables. Second, we propose BayesDisc as a practical algorithm that effectively mitigates non-convexity, only requiring standard deep learning optimizers. Finally, we demonstrate on several datasets that BayesDisc successfully improves causal variable identification over InfoMax SSL baselines.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Causal Representation Learning, Identifiability Guarantee, Self-Supervised Learning, InfoMax Criterion, Discernment Loss

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_RP1KOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This study proposes a new framework for causal representation learning (CRL) in a multi-view setting. Unlike existing multi-view CRL methods, it addresses a setting in which group assignments are unobserved, random, and sample-dependent. The study shows that the latent variables are group-identifiable if the encoder achieves the proposed bilevel Disc-InfoMax objective. It also proposes a practical estimation framework for learning shared latent variables and the corresponding mask simultaneously while avoiding feature collapse through Bayesian mask inference and bandwidth annealing of the gating function. The numerical experiments show that the proposed framework outperforms non-multi-view baselines and achieves performance close to that of a multi-view baseline with true grouping labels in some settings.

Critical strengths
  1. This study proposes a novel condition for (block) identifiability in multi-view CRL that does not require known grouping labels. Since grouping indices are not given in advance in many practical cases, the framework could be useful in such settings.

  2. This study provides a practical estimation framework for learning an encoder from paired observations by minimizing their alignment mismatch. A gating mechanism (proximity mask) focuses on shared latent variables while avoiding feature collapse.

Critical weaknesses
  1. This study does not provide a consistency guarantee for the estimation framework.

  2. Related to the above, the estimation framework relies on inner–outer bilevel optimization, which seems unstable and does not scale well to higher-dimensional latent variables (Fig. 3 and Table 8). This raises questions about the applicability of the framework to more realistic data with potentially larger numbers of latent variables.

Questions for authors

None.

3 comments

Dear reviewer RP1K,

Thank you for your thoughtful review and for considering our work to be well-written and acknowledging our contribution on the novel setup and the corresponding practical algorithm.

Based on your constructive feedback, we have made a revision. All modifications addressing your review are color-coded in orange in the revised paper. Below is our response on each of your points of weaknesses.


1. Statistical Consistency Guarantee for the Empirical Estimator (W1)

To address your concern regarding the consistency guarantee, we have added a formal Statistical Consistency Theorem in Appendix B.3. (Theorem 3). It establishes that the finite-sample empirical estimator consistently converges to the population-level block-identifiable optimum as sample size N→∞N \to \infty. Below is the summary of Appendix B.3.

Theorem 3 (Informal): Under a coupled bandwidth schedule satisfying lim⁡N→∞σN=0+\lim_{N \to \infty} \sigma_N = 0^+ and lim⁡N→∞σNN=∞\lim_{N \to \infty} \sigma_N \sqrt{N} = \infty, the empirical minimizer θ^N∈arg,min⁡θ∈ΘL^N(gθ,t;σN)\hat{\theta}_N \in \underset{\theta \in \Theta}{\operatorname{arg\\,min}} \widehat{\mathcal{L}}_N(g_\theta, t; \sigma_N) satisfies:

dist(θ^N,Θ∗_0+)→p0as N→∞,\mathrm{dist}(\hat{\theta}_N, \Theta^*\_{0^+}) \xrightarrow{p} 0 \quad \text{as } N \to \infty,

and the limiting encoder induced by any accumulation point block-identifies the shared causal variables CS\mathbf{C}_S.

Proof Sketch (Appendix B.3):

  1. Uniform Boundedness & Lipschitz Regularity: For any σ>0\sigma > 0, the per-coordinate objective vσ(δ)=(Φj(σ)(δ)δ)2−λΦj(σ)(δ)v_\sigma(\delta) = (\Phi_j^{(\sigma)}(\delta)\delta)^2 - \lambda \Phi_j^{(\sigma)}(\delta) is uniformly bounded on R\mathbb{R}, and the parameter mapping θ↦∑j=1mvσ(Δj(θ))\theta \mapsto \sum_{j=1}^m v_\sigma(\Delta_j(\theta)) is O(σ−1)\mathcal{O}(\sigma^{-1})-Lipschitz on compact Θ\Theta.
  2. Uniform Convergence via Coupled Rate: By metric entropy covering-number bounds on compact Θ\Theta, the empirical risk error scales as Op((σNN)−1)\mathcal{O}_p\big((\sigma_N \sqrt{N})^{-1}\big). With the coupled rate σNN→∞\sigma_N \sqrt{N} \to \infty and a uniformly consistent differential entropy estimator, the empirical risk converges uniformly in probability: sup⁡θ∈Θ∣L^N−Ljoint∣→p0\sup_{\theta \in \Theta} |\widehat{\mathcal{L}}_N - \mathcal{L}_{\text{joint}}| \xrightarrow{p} 0.
  3. Sandwich Trapping (M-Estimation): Evaluating θ^N\hat{\theta}_N against the ideal population minimizer θideal∈Θ∗_0+\theta^{\text{ideal}} \in \Theta^*\_{0^+} traps the empirical risk between −λ∣S∣0−ε(σN)-\lambda \\|\mathbf{S}\\|_0 - \varepsilon(\sigma_N) and −λ∣S∣0+2ηN-\lambda \\|\mathbf{S}\\|_0 + 2\eta_N. As N→∞N \to \infty, both bounds collapse to −λ∣S∣0-\lambda \\|\mathbf{S}\\|_0. By Wald's consistency for M-estimators on compact Θ\Theta, dist(θ^N,Θ∗_0+)→p0\mathrm{dist}(\hat{\theta}_N, \Theta^*\_{0^+}) \xrightarrow{p} 0. □\square

2. High-Dimensional Scaling & Dataset Difficulty (W2)

We clarify that the lower metrics observed at n=16n=16 and n=32n=32 reflect the difficulty of the synthetic validation benchmark, rather than an intrinsic failure of BayesDisc in higher dimensions:

  • BayesDisc Consistently Outperforms Strong SSL Baselines
    • In Table 1 (n=16n=16) and Table 10 (originally Table 8 in the initial submission; n=32n=32), BayesDisc outperforms InfoNCE and LeJEPA in both MCC and R2R^2.
    • Because standard SSL methods (InfoNCE and LeJEPA) are wisely established to scale effectively to higher latent dimensions in practical applications, their severe performance drop here confirms that the degradation is driven by the difficulty of this particular benchmark.
    • As detailed in Appendix G.2.2, even the supervised True Mask oracle drops MCC and R2R^2.
  • Theoretical Verification Setup vs. Realistic Scenarios:
    • This synthetic dataset was deliberately designed to stress-test and validate the theoretical results, not to represent realistic self-supervised learning pipelines.
    • For realistic scenarios, we have provided the results on Causal3DIdent benchmark. (Section 4.2)

We hope to have clarified most of your stated concerns. Please let us know if there are any additional information we can provide to further clarify our contributions. Thank you once again for your time and effort for reviewing our paper!

Kindly,
Authors

Dear Reviewer,

We would like to thank you again for the insightful review. We have provided a detailed answer in the rebuttal as well as the modified paper. We hope to have resolved all your concerns and will be happy to discuss further if any concerns do remain.

Regards,

Authors

I appreciate the authors' rebuttal. The added consistency theorem addresses my concern in principle. However, the assumptions and conditions required for the consistency guarantee should be stated and discussed more clearly in the main text so that readers can assess the scope of the result. It would also be helpful if the authors provided a more detailed and accessible proof, as the current presentation makes it difficult to follow the argument. Overall, I maintain my score.

Reviewer_XsxmOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Disc-InfoMax for learning invariant causal representations without explicit view labels. A proximity mask selects coordinates for the discernment term, while BayesDisc combines a Bayesian proximity mask, bandwidth annealing, and SIGReg for practical training. The paper claims block-identifiability and evaluates the method on numerical datasets and Causal3DIdent.

Critical strengths

1.Inferring shared content from paired representations is a useful direction, and BayesDisc is simple to integrate into an InfoMax SSL pipeline.

2.The reported MCC gains are substantial on the numerical benchmarks. Sixteen independent runs, OOD evaluations, and downstream causal discovery provide useful empirical evidence.

Critical weaknesses

1.The central identifiability claim is not established. Appendix B.1 treats the limit of finite-bandwidth bi-level optima as though it optimized the limiting regularizer. An invertible mixing and rescaling can instead win the inner objective without block-identifying the shared variable. This review gives a concrete counterexample:

Take paired masked latents (U, V₁) and (U, V₂), where U, V₁, and V₂ are independent uniform variables on the unit interval, and the mixing function is the identity. Only the first causal coordinate is shared; the second is replaced by nuisance noise in both views. The masked-out causal variable can itself have an admissible nonlinear mechanism. Consider a compact encoder class containing the identity and the invertible map h(z₁,z₂) = (z₁ + 10z₂, 10z₂). Both admit auxiliary projections to the uniform unit square with bounded Jacobian determinant, so both entropy terms can be zero. Neither coordinate of h alone identifies U.

Use the admissible proximity mask Φσ(δ) = exp(−δ²/(2σ²)). At bandwidth 0.01, direct integration of the discernment term gives approximately 8.7623 × 10⁻⁷ for the identity and 1.7705 × 10⁻⁷ for h. The ratio tends to 1/5 as the bandwidth tends to zero. Thus h is the unique inner minimizer in this class at sufficiently small bandwidths; the outer optimization cannot select the excluded identity encoder. This exposes the distinction between the limit of the stated finite-bandwidth bi-level optima and optimizing the limiting regularizer over a larger set. It concerns Theorem 1 under the stated encoder-class assumptions, not a claim that BayesDisc fails on every empirical dataset.

2.The theory does not cover the implemented training procedure. Theorem 2 assumes global minimizers of a joint objective containing the outer regularizer. Algorithm 1 uses a detached mask and annealing without that explicit term. The Gaussian residual model is also not implied by Gaussian representation marginals.

3.The experimental setting is narrower than standard SSL. Numerical pairs use controlled observability patterns, and Causal3DIdent pairs use latent-space nearest-neighbor retrieval. These results support learning from structured pairs, but do not establish the broader claim for ordinary image augmentations.

Questions for authors

1.How does Theorem 1 exclude the two-dimensional counterexample in the verification note below? Which additional assumption would make the finite-bandwidth optimization argument valid?

2.What guarantee remains for the detached-mask BayesDisc update, and can its effect be separated from the bandwidth schedule using the existing experimental setup?

6 comments

Dear reviewer Xsxm,

Thank you for your constructive in-depth feedback of our work and for recognizing the simplicity of BayesDisc implementation and strong empirical results.

Considering your feedback, we have revised our manuscript. All modifications regarding your comments are highlighted in green. Below, we address your concerns and questions point-by-point.


1. Counterexample and Order of Limits in Theorem 1 (W1, Q1)

We appreciate the sharp counterexample regarding the original formulation of Theorem 1. As pointed out, in the initial manuscript, taking the argmin⁡\operatorname{argmin} at finite σ>0\sigma > 0 allowed the O(σ3)O(\sigma^3) tail of the discernment term to act as an unintended tie-breaker against the identity encoder gidg_{\mathrm{id}}.

  • Revision of Theorem 1 and Proof: We have updated Theorem 1 and Appendix B.1 so that the limit σ→0+\sigma \to 0^+ is evaluated first on the objectives, followed by the bi-level optimization: Inner (InfoMax):G0+∘≔argmin⁡gθ∈Ginf⁡t∈Tlim⁡σ→0+L(gθ,t;Φ(σ)),\text{Inner (InfoMax):} \quad \mathcal{G}^\circ_{0^+} \coloneqq \underset{{g_{\theta} \in \mathcal{G}}}{\operatorname{argmin}} \inf_{t \in \mathcal{T}} \lim_{\sigma \to 0^+} \mathcal{L}(g_\theta, t; \mathbf{\Phi}^{(\sigma)}),

Outer (Regularization):G0+∗≔argmin⁡gθ∘∈G0+∘lim⁡σ→0+Reg(Φ(σ)).\text{Outer (Regularization):} \quad \mathcal{G}^*_{0^+} \coloneqq \underset{{g_{\theta^\circ} \in \mathcal{G}^\circ_{0^+}}}{\operatorname{argmin}} \lim_{\sigma \to 0^+} \mathrm{Reg}(\mathbf{\Phi}^{(\sigma)}).

  • Resolution: In this limit-first formulation, both gidg_{\mathrm{id}} and the sheared map hh achieve loss 00 in the inner problem (G0+∘\mathcal{G}^\circ_{0^+}). In the outer problem, lim⁡σ→0+Reg(Φ(σ))=−E[∑jI(Δj=0)]\lim_{\sigma \to 0^+} \mathrm{Reg}(\mathbf{\Phi}^{(\sigma)}) = -\mathbb{E}[\sum_j \mathbb{I}(\Delta_j = 0)] strictly selects gidg_{\mathrm{id}} (loss −1-1) over hh (loss 00).
  • Scope: The overall proof sketch remains identical, and Theorem 2 does not require revision because its single-level joint objective Ljoint\mathcal{L}_{\text{joint}} already includes the explicit regularizer λReg\lambda \mathrm{Reg}, which strictly favors invariant alignments at any σ>0\sigma > 0. We updated Section 3.1, Eqs. (6)–(7), and Appendix B.1 accordingly.

2. Theoretical Guarantee of BayesDisc & the Role of stop_gradient (W2, Q2)

BayesDisc was initially motivated as a practical proxy for Disc-InfoMax. Prompted by your question, we conduted theoretical analysis and discovered that BayesDisc directly inherits the theoretical identifiability guarantee of Disc-InfoMax:

  • Effective Proximity Mask: In revised Section 3.2.2 and Appendix C.5, we show that the detached BayesDisc update is mathematically gradient-equivalent (up to a positive learning-rate scaling γ>0\gamma > 0) to an un-detached joint objective with an effective proximity mask Φ~(σ)\tilde{\mathbf{\Phi}}^{(\sigma)} and regularizer λReg(Φ~(σ))\lambda \mathrm{Reg}(\tilde{\mathbf{\Phi}}^{(\sigma)}).
  • Satisfaction of Proximity Criteria: We prove that for any σ∈(0,2)\sigma \in (0, \sqrt{2}), Φ~(σ)\tilde{\mathbf{\Phi}}^{(\sigma)} strictly satisfies all three deterministic conditions of Definition 1 (Peak Alignment, Monotonic Decay, and Asymptotic Nullification). Thus, the asymptotic block-identifiability guarantees of Theorems 1 and 2 apply directly to BayesDisc.
  • Dropping the Annealing Proxy Concept: Because BayesDisc already holds the formal identifiability guarantee, bandwidth annealing is no longer required as a conceptual proxy for the outer optimization. We revised Section 3.2.2 and Appendix E.2 to frame σ\sigma selection simply as tuning a hyperparameter (either a fixed constant or a schedule) to suit the optimization landscape.

3. The Gaussian Residual Assumption & "Where to End Up" vs. "How to Get There" (W2)

  • Clarifying the Assumption: You are correct that marginal Gaussianity does not formally imply that residuals Δj\Delta_j are jointly Gaussian or distributed as N(0,2.0)\mathcal{N}(0, 2.0). In Section 3.2.1, we originally used the word "induced" (Section 3.2.1); we have corrected this in the text to that we have modeled it in such way. One thing to note is that our identifiability theorems do not require Gaussianity—they depend only on Definition 1.
  • Identifiability Guarantee vs. Optimization Landscape: Identifiability theory characterizes where the global optimum ends up, while training deep networks is about how to get there. Even when an objective possesses global identifiability, optimizing an un-detached proximity mask creates non-convex, sign-reversing gradient regions that trap SGD. BayesDisc was specifically proposed to create a well-behaved optimization landscape within the theoretical guarantee, based on Bayesian foundation.

[continued in the next comment]

4. Direct Implementation of Theorem 2 (W2, Q2)

  • Experimental Results: As reported in Table 9 (Appendix G.1.3), direct optimization still successfully identifies causal variables (MCC≈0.90–0.93\text{MCC} \approx 0.90\text{--}0.93), dramatically outperforming vanilla SSL and validating our theoretical framework. However, BayesDisc achieves strictly superior performance, demonstrating the practical benefit of the detached update.
  • Effect of stop_gradient (W2): Combined with the proof provided in Appendix C.5, the experiment shows that BayesDisc has advantage over the direct Theorem 2 implementation while retaining the same theoretical guarantee.
  • Separating BayesDisc from the Bandwidth Schedule (Q2): The bandwidth σ\sigma is kept identical in BayesDisc and direct Theorem 2 implementation. This confirms that BayesDisc's gains stem directly from the detached gradient dynamics, not from bandwidth annealing.

5. Causal3DIdent & Latent Nearest-Neighbor Retrieval (W3)

  • Benchmark Standard: Causal3DIdent with latent-space nearest-neighbor retrieval is the standard benchmark in multi-view Causal Representation Learning.
  • Why Nearest-Neighbor Retrieval is Essential for Evaluation: In SSL without view labels, an encoder identifies whatever underlying factors are invariant across the paired observations. If standard 2D image augmentations (e.g., cropping, blur) were applied, BayesDisc would identify the factors corresponding to those augmentations—which do not correspond to the 11 physical rendering variables (e.g., 3D object rotation, spotlight position) labeled in Causal3DIdent. This can be thought of as an alternate SCM generating the same augmented images, which does not consist of the variables we aim to identify.

We wish this clarifies your concerns and questions. Please feel free to let us know if you have any additional comments. Thank you once more for your effort on reviewing our work!

Sincerely,
Authors

Dear Reviewer,

We would like to thank you again for the insightful review. We have provided a detailed answer in the rebuttal as well as the modified paper. We hope to have resolved all your concerns and will be happy to discuss further if any concerns do remain.

Regards,

Authors

Reviewer_qKjSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies whether a model can infer shared causal factors and learn corresponding representations without explicit view labels. The authors propose Disc-InfoMax, which selectively aligns paired representations using a proximity mask derived from coordinate-wise residuals. They also introduce BayesDisc, a practical algorithm that computes soft alignment weights from these residuals and uses bandwidth adjustment to control how selectively coordinates are aligned. The paper establishes block-identifiability under specific assumptions and reports improved causal variable identification over the compared baselines on synthetic numerical benchmarks and Causal3DIdent.

Critical strengths
  1. The paper addresses a setting where shared content varies across paired observations and explicit view labels are unavailable. Its selective alignment mechanism uses representation residuals to infer which coordinates encode shared content. Both the problem formulation and the motivation behind the method are clear.

  2. BayesDisc computes the proximity mask in closed form and can be trained with standard deep learning optimizers, avoiding computationally expensive nested bi-level optimization. This makes it straightforward to integrate into existing self-supervised learning pipelines.

  3. The paper provides detailed theoretical proofs, presents its arguments clearly, and includes an extensive set of experiments.

Critical weaknesses
  1. The paper shows improvements over the baselines, but the individual contributions of the proximity mask, stop-gradient operation, and bandwidth annealing remain unclear.

  2. BayesDisc achieves an MCC of 0.963 in the default setting, but this drops to 0.799±0.1500.799 \pm 0.150 at n=16n=16 and further to 0.366 at n=32n=32. These results raise questions about the method’s scalability and reliability in higher-dimensional settings.

  3. An overview diagram showing the BayesDisc training pipeline would help readers understand how its components fit together.

Questions for authors
  1. In the numerical experiments, occluded dimensions are filled with zero with probability 0.5, and active causal variables are deliberately shifted away from zero. How much do the results depend on this zero-filling scheme?

  2. In the overparameterized setting with m=32m=32 and n=11n=11 on Causal3DIdent, the remaining 21 latent dimensions are described as uninformative components. Is this conclusion based mainly on the correlation heatmap, or is there additional evidence supporting it?

2 comments

Dear reviewer qKjS,

We truly appreciate your constructive feedback and for crediting the motivation, formulation, BayesDisc and its experiments.

In response to your feedback, we have made a revision in the manuscript. All modifications addressing your coments are color-coded in blue. Below is our response to each weakness and question from you.


1. Individual Contributions of the Mask, Stop-Gradient, and Bandwidth (W1)

  • Proximity Mask: Dropping the proximity mask (Φ≡1\mathbf{\Phi} \equiv \mathbf{1}) reduces BayesDisc directly to vanilla SSL (specifically LeJEPA, which uses the exact same encoder, SIGReg regularizer, and optimizer). Thus, comparing BayesDisc to LeJEPA in Table 1 is already an exact ablation of the mask.
  • Stop-Gradient Operation: To isolate the contribution of stop_gradient, we implemented a direct, un-detached version of the theoretical objective (Theorem 2) with explicit regularization λReg(Φ)\lambda \mathrm{Reg}(\mathbf{\Phi}) (λ=0.02\lambda = 0.02) and no stop_gradient (Appendix G.1.3, Table 9). The direct theoretical implementation successfully identifies causal variables (MCC ≈0.90–0.93\approx 0.90\text{--}0.93), validating our identifiability theory. However, BayesDisc with stop_gradient achieves superior performance (MCC ≈0.95–0.98\approx 0.95\text{--}0.98). This confirms that detaching Φ\mathbf{\Phi} provides a smoother optimization landscape by eliminating repulsive mask-differentiation gradients. Please note that the results are color-coded as green in the revision since we respond to the reviewer Xsxm with the same experimental results.
  • Bandwidth Selection as an Optimization Curriculum: As proven in revised Section 3.2.2 and Appendix C.5, BayesDisc with stop_gradient is mathematically gradient-equivalent to a single-level joint objective with an effective proximity mask, directly inheriting the theoretical identifiability guarantee. Thus, we reframe bandwidth annealing as bandwidth schduling, and it is no longer an outer-optimization proxy, but an optimization hyperparameter controlling gate selectivity. The σ\sigma values used in the experiments are tuned for the best performance, and deviating from it will most likely worsen the results.

2. Scalability and High-Dimensional Performance (W2)

We received the same concern from the reviwer RP1K. Please refer to the answer #2 of our response to reviwer RP1K.

3. Pipeline Overview Diagram (W3)

In response to your suggestion, we have added an overview diagram (Figure 1) on page 1 of the revised manuscript.

4. Independence from the Zero-Filling and Margin Scheme (Q1)

To directly evaluate the effect of the zero-filling scheme, we conducted a new ablation study and added the results to Appendix G.1.2 (Table 8).

  • Setup: We completely removed the static zero-fill and the sign-shift margin. Occluded coordinates are replaced exclusively by independent marginal resampling.
  • Results: As shown in Table 8, BayesDisc maintains strong identifiability without the zero-gap and zero-fill (Default n=8n=8: MCC 0.948±0.0080.948_{\pm 0.008}, R2R^2 0.978±0.0020.978_{\pm 0.002}), vastly outperforming InfoNCE (0.4410.441) and LeJEPA (0.4470.447) and closely tracking True Mask (0.9730.973).
  • This confirms that BayesDisc does not rely on static zero markers or artificial margins, and separates shared from unshared coordinates purely via empirical residual statistics.

5. Overparameterization and Excess Dimensions on Causal3DIdent (Q2)

To verify the role of the 21 excess latent dimensions when m=32>n=11m=32 > n=11, we compared the correlation matrices across BayesDisc and the baseline LeJEPA:

  • In LeJEPA, representations are entangled across all 32 dimensions, leading to diffuse correlations across the entire latent vector. We have added the correlation matrix of LeJEPA in Figure 5. (Appendix G.3)
  • In contrast, BayesDisc concentrates the 11 true causal factors into the primary matched coordinates. For the 21 excess components, the correlation with the true causal factors are considerably lower than those in the baselines.

We thank you again for your constructive review. We hope these new experiments and clarifications address your concerns, and we welcome any further discussion!

Sincerely,
Authors

Dear Reviewer,

We would like to thank you again for the insightful review. We have provided a detailed answer in the rebuttal as well as the modified paper. We hope to have resolved all your concerns and will be happy to discuss further if any concerns do remain.

Regards,

Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Engaging with the rebuttal

Dear Reviewers,

The discussion phase is coming to an end soon. It will be helpful if you could engage and let us know if we have clarified all your concerns. We have worked very hard to provide the answers to all your questions and indeed the reviews have been very helpful to make our paper stronger.

Let us know if we can alleviate any remaining concerns.

Regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.