acceptodds
Under review as a conference paper at ICLR 2027

Generative Sequence Modeling for Infinite Memory Processes via Predictive States

Abstract

We consider estimating the one-step-ahead conditional distribution of a multivariate stochastic process. Many existing approaches rely on assumptions such as finite-range memory, sparsity, or additivity, which can be poorly suited to processes with long-range nonlinear interactions. However, without such structural assumptions, nonparametric estimation is challenging due to the curse of dimensionality. To address this challenge, we introduce a new estimation approach based on the predictive states of a process, possibly with infinite-range memory. We show that our estimator achieves fast convergence rates when the past history can be compressed into a low-dimensional statistic that is sufficient for predicting the future. Specifically, we show that the statistical complexity of the estimation problem is determined by the intrinsic dimension of the predictive state space. We establish guarantees for an instantiation of our method based on deep neural network estimators, and we support these theoretical results with experiments.

Primary area:
learning on time series and dynamical systems
Keywords:
time series, stochastic processes, dynamical systems, generative modeling, predictive state representations, deep neural networks, conditional distribution estimation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_VGqLOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors propose an estimator for the one-step-ahead conditional distribution of a multivariate stochastic process. Under the presented model assumptions, the authors prove that the predictive states can be injectively embedded via the conditional expectations of k≥2p+1k \geq 2p+1 random Fourier features of future observations, where p is the Hausdorff dimension of the predictive state space. Finally, the authors provide convergence rate guarantees for their estimator using recurrent neural networks and mixture density networks.

Critical strengths
  • Estimating the one-step ahead conditional distribution of a multivariate stochastic process is a relevant problem with many practical applications as the authors highliht in section A.1.
  • Contribution 2 is elegant and provides valuable insights on the connection between the geometry of the state space and the number of features required to represent it.
Critical weaknesses
  • The paper suffer from serious presentation issues. Several pages worth of content are devoted to presenting assumptions; yet, most of them remain undiscussed. Similarly, the results in section 3.4 are not discussed and give no insight on the consequences of the main theoretical resuts. The authors also never explain why the rates they propose are fast. Finally, the authors do not include a related work section in the main text but relegate it to the appendix, making it difficult for the reader to contextualize the contribution. The authors cite a relevant work [1] that similarly employs random Fourier features, but they never explicitly discuss this connection, nor do they compare against it experimentally.

  • The estimator of QZtQ_{Z_t}, despite being the central object of the paper, is never stated explicitly. Likewise, the procedure in Algorithm 1 is generic and omits important mechanisms in line 3 and 4 of the algorithm, I suggest that the authors expand those steps or at least point to the equations in the paper responsible for those steps. Moreover, the experimental section does not verify the theoretical results of theorem 4 and 5. Additionally, the data generating process is fairly simple and low-dimensional, which does not allow to assess the robustness of the estimator in more realistic scenarios.

[1] Downey et al. ; Predictive state recurrent neural networks

Questions for authors
  • Can the authors elaborate on why their rates are to be considered fast?
  • Can the authors state the final estimator explicitly, and clarify which theorem applies to the estimator actually used in practice?
4 comments

Thank you for taking the time to carefully review our initial submission and provide constructive feedback. We have made major revisions to improve our paper based on your comments. We hope that these revisions address your concerns.

The paper suffer from serious presentation issues. Several pages worth of content are devoted to presenting assumptions; yet, most of them remain undiscussed. Similarly, the results in section 3.4 are not discussed and give no insight on the consequences of the main theoretical results. The authors also never explain why the rates they propose are fast. Finally, the authors do not include a related work section in the main text but relegate it to the appendix, making it difficult for the reader to contextualize the contribution. The authors cite a relevant work [1] \textit{Downey et al.; Predictive state recurrent neural networks} that similarly employs random Fourier features, but they never explicitly discuss this connection, nor do they compare against it experimentally.

We agree with all of your points. We made the following changes: (1) we have restructured the paper to improve its readability and flow, (2) we moved the assumptions and theoretical convergence rate results to the appendix, while keeping the parameter estimation procedures for the recurrent neural network and mixture density network in the main text (so this is much clearer now), (3) we have clarified that we meant relatively faster rates are achievable when the predictive state space is lower-dimensional, and (4) we have moved the related work section to the main text which contains sections "Comparisons" and "Purpose of Our Work" to address your concerns about comparisons with \textit{Downey et al.; Predictive state recurrent neural networks} and better contextualizes our contributions. Regarding the insights from our results, we list some here: we clarify the probabilistic objects that are being represented, we explain when these representations based on random Fourier features actually identify the predictive state, we provide estimators and explain what probabilistic objects they are actually converging to, and we give sufficient conditions under which this convergence actually occurs.

The estimator of QZtQ_{Z_t}, despite being the central object of the paper, is never stated explicitly. Likewise, the procedure in Algorithm 1 is generic and omits important mechanisms in line 3 and 4 of the algorithm, I suggest that the authors expand those steps or at least point to the equations in the paper responsible for those steps. Moreover, the experimental section does not verify the theoretical results of theorem 4 and 5. Additionally, the data generating process is fairly simple and low-dimensional, which does not allow to assess the robustness of the estimator in more realistic scenarios.

We agree that the estimator Q^Z^t\hat{Q}_{\hat{Z}_t} of the conditional distribution QZtQ_{Z_t} should have been stated explicitly (previously, it was left implicit based on the definition from the estimated conditional density); please see the end of Section ``Implementation with Deep Neural Networks'' for the explicit estimator. We agree that the algorithm was too generic, so for the sake of concreteness we have added a section directly below Algorithm 1 that presents the details for estimating the parameters of the recurrent neural network and mixture density network. In the experiments section, we have added two higher-dimensional data generating processes and a real data example. Please let us know if you'd like to see additional experiments.

Can the authors elaborate on why their rates are to be considered fast?

By faster we mean that the convergence rate can be improved when the dimension of the predictive state is lower. We have now rewritten all sentences about the speed of the convergence rate to clarify that we meant relatively faster rates can be achieved when the dimension of the predictive state space is lower (i.e., not that the convergence rates are always fast).

Can the authors state the final estimator explicitly, and clarify which theorem applies to the estimator actually used in practice?

Please see the end of Section "Implementation with Deep Neural Networks" for the final estimator. Theorem 5 is about the out-of-sample L2L^2 error based on continuing the fitted recurrence, and the plots in the experiments section show that this L2L^2 error (averaged over 500500 out-of-sample time steps) decreases as the training sample size nn increases.

We are grateful for your thoughtful feedback, and we think the updated manuscript is significantly better. We would appreciate hearing your thoughts, so that we can address any remaining concerns.

Thanks again for your review! We'd appreciate your feedback, and we'd be very happy to address any remaining concerns.

I thank the authors for their rebuttal. The changes are welcome and improve the clarity of the manuscript. I still have several concerns:

  • I agree with the authors that "formalizing ideas and asking fundamental questions is valuable because it can help clarify why certain ideas work", but it remains unclear to me how the proposed estimator and its properties clarify why RNNs, MDNs or other architrectures work. Section 2.4.1 now reads as a description of how a RNN works rather than clarifying why it works.
  • The authors write "our theoretical investigations can also inform the development of architectures for new sequence models based on predictive state representations". Can the authors elaborate on this point? What principles and insights in the work can drive the development of new architectures?
  • I do not agree with listing contributions 3 and 4 while putting the technical content in the appendix. I understand that other reviewers and I raised concerns about clarity due to the dense presentation but if the authors believe they form a core contribution they should rethink their presentation rather than simply moving it to the appendix.
  • The new experiments are welcome, but they are not adequately discussed. For instance, in the C. elegans experiment the authors find it interesting that the k=3 instantiation of their method works best. Why is that the case? Why does performance inversely correlate with k in that experiment? Do the authors find the empirical convergence observed in their synthetic experiments consistent with the theoretical rates in the appendix?

Thank you for your thoughtful response to our rebuttal. We really do appreciate it.

I agree with the authors that "formalizing ideas and asking fundamental questions is valuable because it can help clarify why certain ideas work", but it remains unclear to me how the proposed estimator and its properties clarify why RNNs, MDNs or other architrectures work. Section 2.4.1 now reads as a description of how a RNN works rather than clarifying why it works.

We have updated that sentence to make it clearer that the "certain ideas" we were referring to are the ideas from the literature on predictive state representations (discussed in Sections 1.1.1 and 1.1.2), not RNNs and MDNs. The goal of our work is to explain why learning representations of predictive states works at all (this is not obvious), rather than explaining why RNNs or MDNs work (these are just particular architectures that our ideas can be implemented with). Sections 2.4.1 and 2.4.2 explain how our algorithm is implemented with an RNN and MDN, respectively, because those are the architectures used in our experiments (and supported by the convergence rate results).

The authors write "our theoretical investigations can also inform the development of architectures for new sequence models based on predictive state representations". Can the authors elaborate on this point? What principles and insights in the work can drive the development of new architectures?

We have updated this sentence as well. This is the subject of our ongoing work, so unfortunately we can't give more details about the new architecture, loss functions, etc., or even hints, without giving the ideas away. We really do dislike having to give you this answer here, but we hope you understand that we do have a good reason (and that we're not just being unnecessarily coy or dismissive of your good questions). As a compromise, we have changed that sentence to be more of a general statement.

I do not agree with listing contributions 3 and 4 while putting the technical content in the appendix. I understand that other reviewers and I raised concerns about clarity due to the dense presentation but if the authors believe they form a core contribution they should rethink their presentation rather than simply moving it to the appendix.

We have removed these contributions from the "Our Contributions". We agree that we either had to (1) remove the emphasis on contributions that are now in the appendix, or (2) somehow rework that technical material to fit it in the main text. We decided that it would be better to prioritize the improved exposition of our main ideas, so we went with option (1).

Everything moved to the appendix is now presented as an "additional result" or "bonus result", and everything in the main text is presented as a "main result" (we have added small comments throughout the main text to emphasize this). Our hope is that the convergence rate results are now understood by you and the other reviewers as nice "bonus material" that interested readers can go on and read about in the appendix.

We appreciate your help working with us to improve the exposition of our results. We think it's much more sensible now (main ideas presented clearly in main text, and bonus technical results in appendix for interested readers).

The new experiments are welcome, but they are not adequately discussed. For instance, in the C. elegans experiment the authors find it interesting that the k=3 instantiation of their method works best. Why is that the case? Why does performance inversely correlate with k in that experiment? Do the authors find the empirical convergence observed in their synthetic experiments consistent with the theoretical rates in the appendix?

We have added more discussion about the C. elegans experiment (see "Our interpretation"). For example, we discuss the interpretation that our estimator takes advantage of intrinsic structure in the C. elegans data which allows the past information to be compressed into a lower-dimensional representation (OTOH, the baseline vanilla MDN estimator is not able to take advantage of this structure).

Yes, the empirical convergence rates are consistent with our theoretical convergence rates. Of course, there are many unknown constants, so it's tough to say anything more rigorous than this (certainly, if the estimation error was constant or increasing with the sample size, then no, that would not be consistent with our theoretical convergence rate results).

We believe the paper is much better now after incorporating your second round of feedback.

Please let us know if you have any additional concerns.

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

The paper considers estimating the one-step-ahead conditional distribution of a multivariate stochastic process based on the predictive states of the process. They showed theoretically that the predictive states can be identified from the conditional expectations of a finite number of random Fourier features of a finite block of future observations. RNN and MDN are used in practice to conduct the estimation on a toy example.

Critical strengths
  1. The motivation is clear and the story is self-contained.
  2. The proof is complete and it's good that they breakdown the theoretical results in parts and explain what each piece is used for in Section 1.1.
Critical weaknesses
  1. Only low-dimensional synthetic experiments are considered.
  2. Why would people from broad machine learning and deep learning community be interested in this work when Deep SSMs (e.g., S4, Mamba) are proposed to address the exact same problem while being able to scale up to high dimension and to large real-world datasets? Specifically in the context of stochastic process, [1] proposed a Gaussian process that compresses history into finite evolving states, based on SSM, for online predictions. The paper ignores this line of related works completely.
  3. The presentation and notation is a bit too dense. A lot of math notations and terms (e.g., filtration, "predictively sufficient") are thrown in the text without definitions. Section 3 lists 11 math-heavy assumptions without any intuition and practical insights. It feels like a compressed statistic journal submission rather than a submission for ML conference.

[1] Chen et al. Recurrent Memory for Online Interdomain Gaussian Processes. NeurIPS 2025

Questions for authors

See weakness.

4 comments

Thank you for carefully reading our initial submission, and for the time and effort you have put into your detailed feedback. We have made major revisions to the paper based on your comments. We sincerely hope that these changes have addressed your concerns.

Only low-dimensional synthetic experiments are considered.

We agree that the experiments were limited. To address this concern, we have added two higher-dimensional data generating processes and a real data example. We welcome further suggestions, please let us know.

Why would people from broad machine learning and deep learning community be interested in this work when Deep SSMs (e.g., S4, Mamba) are proposed to address the exact same problem while being able to scale up to high dimension and to large real-world datasets? Specifically in the context of stochastic process, [1] \textit{Chen et al. Recurrent Memory for Online Interdomain Gaussian Processes. NeurIPS 2025} proposed a Gaussian process that compresses history into finite evolving states, based on SSM, for online predictions. The paper ignores this line of related works completely.

First, we believe our initial submission lacked clarity, and that this led to a misunderstanding about the purpose of our work. Since we believe other readers may also have this question, we have added an entire sections "Comparisons" and "Purpose of Our Work" to address your question as to why the ML community should care. We hope that this answers your question.

Second, we disagree with the characterization that our approach cannot scale to high-dimensional datasets and large real-world datasets. This might have been true a few years ago before people starting revisiting scaling and parallelizing RNNs. We have cited recent work (from the last two years) that scales nonlinear recurrent neural networks to massive datasets. There is no reason that our approach (when implemented with an RNN) cannot also scale in this way.

Lastly, in the "Comparisons"' and "Purpose of Our Work" sections, we have referenced many works about deep state-space models, including the ones you mentioned. In retrospect, we definitely should have cited the influential SSM papers in the original submission, so thank you for pointing this out.

The presentation and notation is a bit too dense. A lot of math notations and terms (e.g., filtration, "predictively sufficient") are thrown in the text without definitions. Section 3 lists 11 math-heavy assumptions without any intuition and practical insights. It feels like a compressed statistic journal submission rather than a submission for ML conference.

We completely agree, it was way too dense. We have reorganized the paper so that it flows much better now. In particular, we have made the following changes: (1) perhaps most importantly, we have moved the technical assumptions and theoretical convergence rate results have been to the appendix, while retaining the parameter estimation procedures for the recurrent neural network and mixture density network in the main text, (2) we have moved the related work section to the main text, (3) we have defined terms (e.g., filtration and predictively sufficient) as requested. Again, we emphasize that we have made major changes to the text to improve its readability.

We are grateful for your feedback, and we believe that the revised version of our manuscript is significantly better. We look forward to hearing your thoughts so we can address any remaining concerns.

Thanks again for your review! We'd appreciate your feedback on our response so we can address any remaining concerns.

I thank the authors for the rebuttal. However, I don't think my concerns are adequately addressed and would keep my score unchanged. Specifically,

We have cited recent work (from the last two years) that scales nonlinear recurrent neural networks to massive datasets. There is no reason that our approach (when implemented with an RNN) cannot also scale in this way.

This is hypothesis, not evidence.

Furthermore, I don't think such a major text revision and rearrangement during rebuttal stage is fair to those who ensure high quality of the manuscript and clarity at the very beginning.

Thank you for your response.

I thank the authors for the rebuttal. However, I don't think my concerns are adequately addressed and would keep my score unchanged. Specifically,

This is hypothesis, not evidence.

You are absolutely right that follow-up work should be done to scale up these models and compare with SSMs like Mamba (we strongly agree, see our conclusion and future work!).

However, doing this properly first requires new parallelization methods for these models. This requires a lot more theoretical development that builds on the ideas from this paper, which could be the subject of many follow-up papers.

Currently, the ML community doesn't have any reason to revisit the older literature on predictive state representations to advance generative sequence modeling. Why would we, when we now have transformers, Mamba, etc.?. Our insights about the faster convergence rates give the ML community new reasons to revisit this literature and develop new parallelization methods to scale up these models. Once this is done, then we can get the empirical evidence that we both want to see (i.e., comparisons with transformers, Mamba, etc. on massive datasets).

For these reasons, we think the ML community would find our paper very interesting.

I don't think my concerns are adequately addressed

Respectfully, could you please clarify a little more about what your remaining concerns are so we can address them? We added higher dimensional experiments, gave reasons why the ML community should care, and improved the flow of the paper, as you requested in your initial response.

Furthermore, I don't think such a major text revision and rearrangement during rebuttal stage is fair to those who ensure high quality of the manuscript and clarity at the very beginning.

We respect your position. We hope you understand that this is not how we see it, though.

Respectfully, the final revised paper is what should be considered in the final decision.

The reviewers requested changes. We revised the paper accordingly, and we responded in the comments with the details of our changes throughout the review process. For example, we moved some assumptions and results to the appendix, and we added more discussions about related work and experiments. As a result, we were able to improve the reading experience. We just wanted to make you happy with our work

Regardless, the revised paper is better because of you, and we thank you for that.

We hope our discussion above about the relevance of our paper to the ML community gives you a reason to reconsider.

Reviewer_r4fEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies generative modeling for potentially infinite-memory stochastic processes through low-dimensional predictive states. The central idea is to represent the predictive state using conditional expectations of random Fourier features of a finite future block. Under geometric and regularity assumptions, the authors show that k≥2p+1k \ge 2p+1 random features are sufficient to identify a predictive-state space of intrinsic dimension (p). In practice, an RNN estimates this representation and an MDN models the next-step conditional distribution.

Critical strengths
  1. The finite dimensional predictive-state identification result is theoretically interesting.

  2. The paper provides a principled alternative to arbitrary RNN hidden states: the representation Z_t is explicitly defined through predictions of future observations.

  3. The statistical analysis connects low-dimensional predictive structure to convergence rates even when the observed process may have infinite-range memory.

  4. The architectural components themselves are standard, while the novelty lies in the predictive-state representation and the associated finite-dimensional identification and convergence theory.

Critical weaknesses
  1. The theoretical guarantees rely on a substantial collection of assumptions, including finite-dimensional predictive sufficiency, finite-horizon identifiability, regular geometry of the state space, smoothness/stability conditions, stationarity, and exponential mixing. It is unclear how broad the practically relevant class satisfying all of these assumptions is.
  2. The empirical evaluation is limited to two synthetic processes and a growing-lag MDN baseline, where the proposed method performs largely comparably rather than demonstrating a clear practical advantage.
Questions for authors
  1. Theorem 2 establishes injectivity of Φ\Phi, but is there any quantitative guarantee on the conditioning of this embedding? In particular, can two substantially different predictive states map to arbitrarily nearby Z representations, and does increasing k improve robustness beyond the 2p+1 identifiability threshold?
  2. How should the proposed predictive-state representation be viewed relative to a state-space-model hidden state, where structure can also be imposed? Is the primary advantage that ZtZ_t has an identifiable predictive interpretation rather than being an arbitrary learned latent representation?
4 comments

Thank you for your careful reading of our initial submission and thoughtful comments. We have made major revisions to the paper based on your feedback. We hope that our changes address your concerns.

The theoretical guarantees rely on a substantial collection of assumptions, including finite-dimensional predictive sufficiency, finite-horizon identifiability, regular geometry of the state space, smoothness/stability conditions, stationarity, and exponential mixing. It is unclear how broad the practically relevant class satisfying all of these assumptions is.

We have added a real data example to show that our ideas practically applicable in real data settings. Some assumptions can, in principle, be checked. For example, it is straightforward to test for stationarity and it is possible to estimate β\beta-mixing coefficients from an observed time series. However, it is not as clear how to test that the other assumptions hold based on an observed time series. To some degree, this is unavoidable, and the assumptions should just be understood as sufficient conditions for the convergence rate guarantees to hold.

The empirical evaluation is limited to two synthetic processes and a growing-lag MDN baseline, where the proposed method performs largely comparably rather than demonstrating a clear practical advantage.

We agree that the experiments were limited in the initial submission. In the experiments section, we added two higher-dimensional data generating processes and the previously mentioned real data example. Please let us know if you'd like to see more simulations or real data examples.

Theorem 2 establishes injectivity of Φ\Phi, but is there any quantitative guarantee on the conditioning of this embedding? In particular, can two substantially different predictive states map to arbitrarily nearby Z representations, and does increasing k improve robustness beyond the 2p+1 identifiability threshold?

This is a fantastic question.

Regarding your first point, yes, under the analyticity conditions we impose on the characteristic function, there exist C,β>0C,\beta >0 such that ∥Φ(s)−Φ(s′)∥≥C∥s−s′∥2β\left\|\Phi(s)-\Phi(s')\right\| \geq C \left\|s-s'\right\|_2^{\beta}. This follows by Lojasiewicz's inequality (see Theorem 6.4 of Semianalytic and subanalytic sets by Edward Bierstone and Pierre D. Milman) applied to f(s,s′)=∥Φ(s)−Φ(s′)∥22f(s,s')=\left\|\Phi(s)-\Phi(s')\right\|_2^2 and g(s,s′)=∥s−s′∥22g(s,s')=\left\|s-s'\right\|_2^2 with M=R2rM=\mathbb{R}^{2r} and K=S×SK=\mathcal{S}\times\mathcal{S}.

Regarding your second point, yes, increasing kk retains this separation guarantee for Φ\Phi. Moreover, increasing kk cannot decrease the Euclidean distance when additional features are used (this just follows by the definition of Euclidean distance), but in practice this comes at the cost of increasing the estimation error and computation cost (possibly by a large amount). In our theoretical analysis, we study a sufficient value of kk needed for this separation guarantee for Φ\Phi, but we suggest choosing the value of kk that performs best in practice via cross-validation. In view of the feedback from the other reviewers, we decided not to include these discussions in the paper to avoid confusion about new technical discussions.

How should the proposed predictive-state representation be viewed relative to a state-space-model hidden state, where structure can also be imposed? Is the primary advantage that ZtZ_t has an identifiable predictive interpretation rather than being an arbitrary learned latent representation?

This is another great question. Yes, one of the main advantages of our approach is that the predictive state representation ZtZ_t is identifiable, estimable, and has a well-defined probabilistic interpretation (i.e., in terms of the conditional distribution of future observations), without even requiring a user-imposed parametric model. In contrast, the hidden state of a state-space model is typically only identified up to invertible transformations. For example, even linear state-space models (let alone nonlinear recurrent neural networks) have identifiability issues: suppose the hidden state evolves as ht=Aht−1+Bxth_t = A h_{t-1} + B x_t and the output is yt=Chty_t = C h_t. Letting Ht=rhtH_t = r h_t for some invertible rr, we have Ht=rAr−1Ht−1+(rB)xtH_t = r A r^{-1} H_{t-1} + (rB) x_t, while yt=Cr−1Ht=Chty_t = C r^{-1} H_t=C h_t. In other words, any coordinate system can be used for the state variable, without changing anything about the outputs yty_t. This identifiability doesn't matter for the purposes of prediction, but it is important if one wants a fixed coordinate system for interpreting the hidden states (e.g., for comparing the predictive states across models, or the same model across training runs). We have added some discussion about this in the ``Comparisons'' section.

We are grateful for your thoughtful feedback, and we think the revised version of our manuscript is significantly better. We are eager to hear your response so that we can address any possible concerns.

Thanks again for your review! We'd really appreciate your feedback on our response.

I appreciate the efforts in the revisions and rebuttals. However, at the current stage, I agree with other reviewers that the presentation issues that still exist. For example, discussions in section 2.4.1 feel especially disconnected from the rest of the papers.

For these reasons, I will maintain my score for now.

Thank you very much for your response.

For example, discussions in section 2.4.1 feel especially disconnected from the rest of the papers.

Thank you for raising your concern about this. It is very important to us that readers understand the purpose of section 2.4.1.

To address the one concern you mentioned, we have added eight additional clarifications throughout the paper that clarify the purpose of section 2.4.1 (and section 2.4.2) and emphasize their connections to the rest of the paper. Please see the updated versions of the following parts of the paper:

  1. The paper outline now explains how every section fits together, with added emphasis on sections 2.4.1 and 2.4.2
  2. The end of section 2.2 now discusses connections with sections 2.4.1 and 2.4.2
  3. The discussion after Algorithm 1 discusses RNN and MDN implementations give in sections 2.4.1 and 2.4.2
  4. The beginning of Section 2.4 outlines the purpose of sections 2.4.1 and 2.4.2
  5. Section 2.4.1 has a better explanation about Z^t\hat{Z}_t being the RNN output, plus links to ZtZ_t from Section 2.2
  6. Section 2.4.2 now reinforces this by linking back to Z^t\hat{Z}_t from the RNN from section 2.4.1
  7. The beginning of Section 3 similarly reinforces that the experiments are with the implementation from sections 2.4.1 and 2.4.2
  8. The conclusion summarizes everything and mentions sections 2.4.1 and 2.4.2 again

It should be much easier for readers to understand the importance of section 2.4.1 now.

I agree with other reviewers that the presentation issues that still exist.

The reviewer who responded so far said, "The changes are welcome and improve the clarity of the manuscript". We believe the point that this reviewer was making about section 2.4.1 was more of a conceptual point (which we have addressed), rather than a point about the presentation.

The other reviewer hasn't responded yet about the revised manuscript, and we have addressed their comments about presentation issues from the initial submission.

Respectfully, could you clarify what you think the remaining presentation issues are? We really do appreciate your time and effort, we just want to make sure we are addressing the correct things (beyond what we have addressed above regarding section 2.4.1).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.