acceptodds
Under review as a conference paper at ICLR 2027

CABLE: Convex Affine Blends of Latent Experts for parametric reduced-order modeling

Abstract

Non-intrusive reduced-order models often learn low-dimensional latent dynamics. However, existing approaches typically trade explicit structure for the flexibility needed to capture parametric and temporal variation. Equation-based models are efficient and analyzable, but can be too rigid for strongly varying dynamics. Neural dynamical models are more expressive but hide the learned vector field inside a network and can be costly to integrate. We introduce CABLE (Convex Affine Blends of Latent Experts), which represents the latent dynamics as a convex mixture of shared affine ordinary differential equation (ODE) experts. A neural gate varies the mixture weights with the parameters, latent state, or time, allowing a single model to adapt across dynamical regimes while retaining explicit operators. These operators enable rollout-error bounds, sufficient conditions for contraction imposed directly on the experts, and a dynamics-based adaptive sampling criterion. Across several parametric systems, CABLE produces accurate and efficient predictions. On a two-parameter Vlasov–Fokker–Planck two-stream instability, uniformly sampled CABLE matches or improves upon parameterized neural ODE, Latent Space Dynamics Identification, U-Net, and Fourier neural operator baselines in typical-case accuracy while providing lower upper-tail errors and the fastest inference. Adaptive sampling further improves upper-tail accuracy, approximately halving the maximum error relative to the best uniformly sampled baseline.

Primary area:
learning on time series and dynamical systems
Keywords:
reduced-order modeling, mixture of experts, scientific machine learning, latent dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_wMPNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents CABLE, an approach for reduced order modeling in a latent space. It is presented as an improvement over parameterized neural ODEs, and differs in the way the latent dynamics are modeled and learned. Instead of using a neural network for the latent dynamics like their PNODE baseline they use a neural network only for a weight function that selects from a set of Affine Experts, that is trained along with this weight function during training and frozen during inference.

Critical strengths

I like the presentation in general: the language is clear and concise, and the figures are visually appealing and not overloaded. I also like the approach of isolating and demonstrating the various model properties through experiments and not just through arguing them.

Critical weaknesses

To me the significant weak point is the experimental section:

  • There is exactly one experimental result where CABLE is numerically compared to Baselines. When comparing Uniform CABLE to PNODE in this table they are within the stated SDs of each other for all but one row, where it is quite close. This makes the abstract sound slightly overstated, even though it is technically correct.
  • It is not clear to me, whether the adaptive sampling that most of the claimed numerical improvement comes from, is actually a property of the CABLE architecture or orthogonal to it, see also my question below.
  • Most of the experimental results are discussed in the appendix, with just a short conclusion on page 7. A comparison to baselines is only done in Table 1 (PNODE, GPLaSDI, U-Net, FNO) and D.2 (just PNODE). So depending on how you look at it there is only one or maybe two datasets / problems with an actual comparison, and in Table 1 the margins are slim, as stated above. This is simply too little to get a conclusive impression of CABLEs performance against the stated baselines.

I believe the work has potential, but as of now the experimental results are just too thin to get a decent picture of CABLEs performance. I would like to see more experiments on different datasets and problems, preferably also problems the other methods were tested on originally (because that usually means they are well tuned on those). Depending on what those experiments show, I see to paths forward: either the pure numerical performance against the baselines already justify the approach, or if the margins stay as slim, the additional properties (error bounds, stability by construction, …) could justify the approach. In this case I would like a more detailed discussion and demonstration of those properties especially compared against the baseline. Of course a negative result (no significant improvement over baselines) could also be interesting to the community, but for a conclusion like that more experiments are needed as well.

Questions for authors
  • Paper and appendix mention error bounds and stability by construction as advantages. The experimental section refers most of the discussion on this to the appendix. I am missing a discussion of these properties for the baselines. E.g. is enforcing stability structurally impossible for the Baselines?
  • l. 326-327: „For CABLE, we leverage the fact that our model defines a latent vector field … at every training and candidate input parameter.“ -> would this not also apply to other baselines (at least PNODEs)? Why is there no adaptive-PNODE column in Table 1?
  • Introduction mentions „parameter studies, uncertainty quantification, and optimization.“ as potential use-cases. Are there benefits to the CABLE in any of those apart from inference time? Maybe the convex blend of experts has advantages against a pure Neural ODE when it comes to using this model in an optimization problem?
3 comments

We thank the reviewer for their comments and concrete suggestions. We believe that addressing the points raised in this review has substantially improved the empirical argument of the manuscript.

Adaptive PNODE (Section 6.3, Table 2, Appendix B.3). The reviewer is correct that our sampling criterion can be applied to any dynamics model with a learned vector field. We added adaptive PNODE with the same 49-trajectory budget applied to CABLE (Table 2). PNODE's 99th-percentile error decreases from 42±6%42\pm6\% to 37±3%37\pm3\%, while maximum error changes from 68±9%68\pm9\% to 72±17%72\pm17\%, with no clear improvement. CABLE's improvement is much larger from adaptive sampling, as its maximum error decreases from 50±9%50\pm9\% to 26±3%26\pm3\%. Figures 3 and 8 show that CABLE, unlike PNODE, concentrates acquisitions in the difficult low-temperature region, leading to the lower tail errors.

Additional benchmarks (Section 6.2, Table 1, Appendix B.2). On the published 1D Burgers' benchmark from PNODE's manuscript, CABLE matches PNODE's accuracy (1.11±0.52%1.11\pm0.52\% versus 1.04±0.69%1.04\pm0.69\% mean error) at roughly half the inference time (52.1 ms versus 110.4 ms). On the published 2D benchmark from GPLaSDI's manuscript, all three ROMs are tested with adaptive sampling. GPLASDI is the most accurate and fastest model, but CABLE and GPLaSDI achieve mean errors below 1%. CABLE's inference time is still quite fast 19.3 ms compared to GPLaSDI's 4.6 ms.

Stability and error analysis (Section 4). Stability is not exclusive to CABLE, and stable neural ODEs exist, which we now cite and discuss. However, the existing approaches we are aware of require learning a Lyapunov function or energy function, and require evaluating that function’s gradient during rollout. CABLE instead imposes stability directly on its expert matrices, with an explicit learned metric PP. This makes every mixture of experts contractive, and adds no neural-network or gradient evaluations during prediction.

Optimization and uncertainty quantification. We have not tested these applications directly, but it remains an interesting future direction. We note that gradient-free optimization methods such as Bayesian optimization or differential evolution often only require a large number of model runs, so any advantage in inference time is valuable for those problems. Explicit operators may support sensitivity analysis or constrained surrogate design, but these applications remain future work.

My biggest concerns were addressed:

  • The authors added an ablation on the adaptive sampling scheme to PNODEs as well.
  • The experimental section was expanded with two more problems, where Cable is compared against PNODE and GPlasDI

I still have some objections: p6l323 „These approaches guarantee stability around a single equilibrium, and require evaluating the learned function’s gradient during rollout. CABLE instead imposes stability directly on its expert matrices, with an explicit learned metric P.“ I think (have not checked) stability is also enforceable structurally through weight limits, in this case no explicit function has to be learned. More importantly: is the explicit learned metric P of cable not also a Lyapunov function?

p17l894 „ All models train for 20,000 epochs, with CABLE/PNODE using the derivative-matching loss while GPLaSDI uses its own model loss.“ If I recall correctly the PNODE paper uses solver based training. The original Vlasov-Fokker-Planck experiment uses both derivative and solver based training if I recall correctly. I agree that using similar training strategies for CABLE and PNODE makes sense for a fair comparison, but the authors should clarify why they chose to omit the solver training for those examples. An ablation that shows, that this training regime doesn’t unfairly favor one of the methods would also be good.

We thank the reviewer for their follow-up and are glad that our revisions addressed the main concerns. We address the remaining objections below.

Stability (Section 4). We agree that our wording was too strong. CABLE's learned PP is a Lyapunov function, and stability can also be enforced structurally in neural ODEs. We removed the mention of gradient evaluations (end of section 4) and clarify that CABLE uses the established common-Lyapunov construction as an optional constraint on its expert library. We do not claim that the common-Lyapunov construction itself is exclusive to CABLE.

PNODE training (Appendix B.2). The reviewer is correct that the original PNODE uses solver-based training. For the added Burgers benchmarks, we chose derivative matching for both CABLE and PNODE to provide a common training procedure and avoid the additional cost of differentiating through ODE solves during joint autoencoder/dynamics training. Derivative matching is also commonly used in equation-based ROM approaches, including Champion et al. (2019) and LaSDI variants. We have clarified this choice in Appendix B.2.

We note that PNODE and CABLE do receive short-horizon rollout fine-tuning on the Vlasov problem (Appendices A and B.3), with gradients detached between steps. Unfortunately, we do not think we will be able to run solver-based training ablations before the end of the rebuttal period. We instead clarify our training choices and explicitly acknowledge that we use a different training approach than the original one the Burgers' problems.

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

This paper proposes a reduced-order model for predicting how parameterized dynamical systems evolve called CABLE. CABLE uses a mixture-of-experts architecture, where each expert is an affine ordinary differential equation. The main gating mechanism is a neural model which takes as input time along with any other known quantities, and they also provide the option of including the latent space in the form of a state-dependent gate. The authors use this explicit expert structure to derive a rollout-error bound, impose stability conditions, and select new simulations through adaptive sampling. Their main comparison is on a two-parameter Vlasov–Fokker–Planck two-stream instability. On this benchmark, they find that CABLE performs competitively with existing methods, while improving on high-percentile errors along with reducing inference time. They further show that adaptive sampling reduces both the magnitude and variability of tail errors on this benchmark. In addition to this comparison, they also conduct smaller experiments to isolate specific capabilities. These include a Van der Pol oscillator to test time extrapolation and a viscous Burgers problem to test adaptive sampling

Critical strengths

One of the strongest contributions of CABLE is its ability to change along a trajectory. The model can represent systems whose dynamics are extremely complicated and change within a single trajectory while keeping explicit affine experts. The Vlasov–Fokker–Planck benchmark represents one of these very difficult problems and their method handles it well, especially when concerning the tail of the error distribution. This is consequential because it speaks to a rather unique use case that CABLE addresses that other methods may not.

Another strength is the clear inference speed advantage. Its scheduling gate lets the mixture weights be precomputed before rollout, speeding up the process. This is consequential because this speed could be useful when the model must make repeated predictions across many parameter settings.

Critical weaknesses

One weakness present in the paper is that the empirical evidence is concentrated in one main benchmark. The main evaluation across baselines is only on one simulated system which may not be sufficient to demonstrate CABLE’s robustness and adaptability to other domains. This is combined with the fact that the accuracy advantage is restricted to primarily hard cases and not more general results. This is still relevant in settings where avoiding large failures is important however reduces the strength of the overall contributions.

The validation of the contribution of adaptive sampling is also a bit thin. First, the paper compares CABLE with adaptive sampling only with other methods using uniform sampling which may obstruct the clearness of the results. Second, within the CABLE method the paper compares adaptive and uniform sampling but does not include any other sampling strategies. Comparing with even other simple sampling methods would better illuminate how strong of a contribution the sampling is.

Questions for authors

In table 1, were all the other applicable methods based on uniform sampling or were any competing methods with other types of sampling used?

1 comment

We thank the reviewer for their feedback. We agree that broader evaluation was required to assess how CABLE compared to baselines. We have added two examples based on the reviewer's feedback, along with a sampling ablation.

Broader evaluation (Section 6.2, Table 1, Appendix B.2). We added benchmarks from the original PNODE and GPLaSDI papers (see the general response point 2).

On the forced 1D inviscid Burgers' problem from the initial PNODE manuscript, CABLE and PNODE have similar peak errors at 3.97±0.60%3.97\pm0.60\% and 4.02±0.16%4.02\pm0.16\%, while CABLE halves inference time (52.1 versus 110.4 ms). On a 2D viscous Burgers' from the original GPLaSDI manuscript, all models are accurate. GPLaSDI is the most accurate and fastest model, but CABLE and GPLaSDI both achieve mean errors below 1%. CABLE's inference time is still quite fast 19.3 ms compared to GPLaSDI's 4.6 ms.

Sampling comparisons (Section 6.3, Tables 2 and 7). In our original Table 1 (now Table 2), all non-CABLE baselines used uniform sampling. Revised Table 2 adds adaptive PNODE with the same acquisition criterion and 49-trajectory budget as CABLE. PNODE's 99th-percentile error improves from 42±6%42\pm6\% to 37±3%37\pm3\%, without any clear maximum-error improvement. This provides significantly less benefit than adaptive CABLE, which halves maximum error to 26±3%26\pm3\%. Figure 8 shows adaptive PNODE focuses less samples in the difficult low-temperature region, which accounts for the large errors compared to adaptive CABLE. Table 7 (Appendix B.3.1) now compares adaptive, uniform, and random sampling with CABLE. Maximum errors are 26±3%26\pm3\%, 50±9%50\pm9\%, and 53±13%53\pm13\%, respectively, showing our adaptive sampling criterion performs best of the three methods.

Reviewer_s9YgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CABLE, a non-intrusive parametric reduced-order model in which the latent state of an autoencoder follows ż = Σ_k w_k (A_k z + b_k), a convex combination of K shared affine ODE experts weighted by a softmax MLP gate that sees time and the PDE parameters (in one example also the latent state). The paper gives a Gronwall-type rollout error bound, a common-Lyapunov stability condition with a stable-by-construction parameterization, and an adaptive sampling criterion (15) based on disagreement of the learned vector field across parameters. Experiments cover a forced heat equation, Van der Pol, Burgers, heat relaxation and a two-parameter Vlasov–Fokker–Planck two-stream instability against PNODE, GPLaSDI, U-Net and FNO.

Critical strengths

The Vlasov benchmark is hard and well instrumented: 441 parameter cells, 5 seeds, matched dynamics-parameter counts against PNODE, and tail percentiles rather than only means. The convexity ablation (Table 5) is the one cleanly isolated finding: with a linear last gate layer, the maximum and 99th-percentile errors under uniform sampling rise from 50 to 66 and from 31 to 49, so the simplex constraint is what buys the tail accuracy. The theory is presented honestly: Theorem 2 is labelled a restatement, and Fig. 9 shows the logarithmic-norm bounds growing exponentially on a held-out Vlasov trajectory.

Critical weaknesses

My main concern is positioning. A convex blend of affine systems with a learned scheduling gate is a Takagi–Sugeno model, the trajectory-piecewise-linear ROM of Rewienski and White [1], and deep LPV identification with a learned scheduling map [2]; DVBF-LL, which the paper does cite, has the same transition, including softmax weights in its implementation (App. B there). The paper cites LPV control and PWA systems but not these works, the explicit-operator non-intrusive literature (operator inference, Koopman autoencoders) is absent, and Contribution 1 presents the model class as new. This turns the paper into an application of a known class, which must then stand on its experiments. I also note that at fixed µ a scheduling-gate CABLE is a linear time-varying affine ODE whose µ-dependence lives entirely in the gate MLP (by my count about 30k of the 33k dynamics parameters in Table 4), which the paper never states.

Second, the experiments establish less than the abstract's "matches or improves ... typical-case accuracy". Uniform CABLE's mean and median (5.6/4.7) are no better than U-Net's (5.4/4.5), and its maximum error ties FNO (50 ± 9 vs 50 ± 2); the gain is at the 90th–99th percentiles. The headline adaptive result (maximum 50 → 26) is confounded: the adaptive run's autoencoder has peak error 8.8% against 27.9% for the uniform one (Table 1), so no uniform latent ROM could get below about 28%, and the Burgers experiment froze the autoencoder precisely to avoid this. No alternative acquisition rule is run (random, or GP variance), so criterion (15) is not isolated. Two baseline settings also look unfavourable, though I am not familiar enough with ROM benchmarking to weigh them: U-Net and FNO appear to have hit the five-hour cap at under 1000 epochs (Table 4), and GPLaSDI is run at N_z = 20, interpolating 420 coefficients per trajectory.

[1] Rewienski and White. A trajectory piecewise-linear approach to model order reduction and fast simulation of nonlinear circuits and micromachined devices. IEEE TCAD 2003. [2] Verhoek et al. Deep-Learning-Based Identification of LPV Models for Nonlinear Systems. IEEE CDC 2022.

Questions for authors
  1. Please repeat the Vlasov adaptive-vs-uniform comparison with the autoencoder frozen to the uniform one, as for Burgers. How much of the 50 → 26 reduction survives?
  2. Please compare criterion (15) with at least one alternative rule (random or GP variance) at the same 49-trajectory budget.
  3. Were U-Net and FNO converged at the cap, and what does GPLaSDI give at the latent dimension its authors use, with its own sampling?
  4. Both adaptive columns of Table 5 report identical autoencoder error 8.8 ± 0.1. Did the unconstrained run perform its own acquisition or reuse the softmax run's trajectories?
  5. How does CABLE differ from TPWL [1], Takagi–Sugeno models, DVBF-LL and LPV identification [2], and why are operator inference and Koopman autoencoders not baselines?
4 comments

We thank the reviewer for the careful review. We first answer the questions in order.

1. Frozen autoencoder (Table 7, Appendix B.3.1). As requested, freezing the uniformly trained autoencoders gives maximum errors of 65±19%65\pm19\% for adaptive sampling and 55±4%55\pm4\% for uniform sampling. However, the uniformly-trained autoencoder has high reconstruction error (27.9±1.527.9\pm1.5), and we should not expect the dynamics model to achieve much higher accuracy than the autoencoder. Thus, the adaptive criterion can sample points that are poorly represented by the autoencoder, but we are limited by the fact that we cannot improve our autoencoder representation at these points if it is frozen.

For this reason, we believe the stronger experiment is to freeze the adaptively-trained autoencoders, which achieve much lower peak reconstruction errors of 8.8±0.1%8.8\pm0.1\%. Freezing this more accurate autoencoder, adaptive dynamics training then reduces maximum error from 58.9±19.2%58.9\pm19.2\% with uniform sampling to 37.7±15.0%37.7\pm15.0\%. The 99th-percentile error also reduces from 42.9±17.4%42.9\pm17.4\% to 23.0±10.7%23.0\pm10.7\%, while mean and median prediction accuracy remain similar. We note that there is some overlap of accuracy ranges for both sampling strategies. However, on each seed we saw that adaptive sampling was more accurate than uniform in peak/tail errors, while uniform sampling was slightly better in median prediction accuracy. Thus, our adaptive sampling criterion improves joint training and can also improve dynamics with an accurate autoencoder held fixed (Discussed in Section 6.3).

2. Alternative acquisition (Table 7, Appendix B.3.1). We now also test random acquisition using the same setup and training procedure as the initial test of our adaptive sampling. With joint training and the same 49-trajectory budget, adaptive, uniform, and random sampling achieve maximum errors of 26±3%26\pm3\%, 50±9%50\pm9\%, and 53±13%53\pm13\%, respectively. Random sampling accuracy is comparable to uniform sampling, with greater variability. Our proposed criterion remains the most accurate sampling approach we tested.

3. Baseline training. The U-Net loss had plateaued at the five-hour cap. The FNO loss was still decreasing slowly. Extending FNO to 2,000 epochs (which requires slightly over double the training time) reduced overall mean error from 13.1±0.9%13.1\pm0.9\% to 10.7±0.9%10.7\pm0.9\%, while maximum error slightly increased from 50±2%50\pm2\% to 52±3%52\pm3\%. However, the 5 hour training-time cap already gave the FNO nearly 20% more training time compared to the PNODE, which was the most expensive ROM baselines. The extended run also does not change our conclusions as FNO’s mean error (10.7%) remains roughly twice that of CABLE, and its maximum error is almost unchanged. We therefore prefer to keep the original experiment in the paper, if the reviewer agrees.

The Vlasov ROMs use matched 20-dimensional latent spaces. We have not tested lower dimensions for GPLaSDI on Vlasov. We now note that GPLaSDI's error exceeds 100%100\% on some training trajectories (Section 6.3), and so the model struggles even on training data due to the affine latent dynamics restriction. At training points, there is essentially no GP error as the GPs return the fitted coefficients, so the error is in representation of the latent dynamics, not interpolation.

On the new 2D Burgers' benchmark, GPLaSDI uses a five-dimensional latent space and its own GP-based acquisition. GPLaSDI is both most accurate and fastest on this problem, with 0.29±0.07%0.29\pm0.07\% mean error versus CABLE's 0.77±0.24%0.77\pm0.24\%, and 4.6 ms inference versus 19.3 ms for CABLE (Table 1, Appendix B.2).

4. Unconstrained gate (Table 6, Appendix B.3.1). The unconstrained run performed its own acquisition and joint training. The peak autoencoder errors are 8.790±0.137%8.790\pm0.137\% for CABLE and 8.786±0.089%8.786\pm0.089\% for the unconstrained gate, but both were rounded to 8.8±0.1%8.8\pm0.1\%.

5. Prior work (Section 1). We agree that convex affine mixtures and common-Lyapunov constructions are established. However, the mentioned methods are different either in model class, training, or application. In particular, none of the listed references tackle non-intrusive surrogate modeling for parametrized PDEs, which is the main focus of CABLE. More specifically, TPWL is intrusive and constructs operators from full-order equations while CABLE is nonintrusive and learns from snapshots. DVBF-LL learns discrete-time probabilistic transitions, and deep LPV identification learns a scheduling map from input–output data. CABLE jointly learns nonlinear compression and deterministic continuous-time dynamics for parametric reduced-order modeling, with rollout analysis and adaptive training (Sections 3–5). We have revised our related works and Contribution 1 based on these new references.

(continued...)

Operator inference and Koopman autoencoders are relevant and now cited. We selected PNODE and GPLaSDI as established parametric ROM baselines for neural and explicit equation-based latent dynamics. We note that CABLE with K=1K=1 is essentially a Koopman-style autoencoder and this fails on the forced heat problem from appendix B.1. Op-inf, to our knowledge, also generally uses linear POD compression which is known to struggle with advection or filamentation-dominated problems like our Vlasov example.

We address some of the reviewers other points below.

Dependence on the neural gate (Section 4). The reviewer is correct that most of our dynamics model parameters live in the neural gate, along with the μ\mu-dependence, and this choice is intentional. For scheduling gates, the weights can be precomputed, and the expert-based error/contraction bounds do not depend on the gate's size or complexity. For state-dependent gates, the bounds remain valid but depend on the Lipschitz constant of the gate, as stated in Theorems 1 and 2.

Accuracy claims (Section 6.3, Table 2). We now emphasize competitive typical-case accuracy and lower 90th–99th percentile errors among uniformly sampled models, not universal superiority. Uniform CABLE's maximum error ties FNO, and we now have evidence that adaptive improvement holds for both joint training and a frozen autoencoder, provided the autoencoder is accurate.

Thanks for the extra ablations and for fixing the positioning. The adaptive result did not survive the test I asked for, so I keep my score.

We thank the reviewer for taking the time to consider our revision. We would like to clarify two final points:

1. Our primary focus is the joint training of the autoencoder and latent dynamics, as is common in related approaches such as Champion et al. (2019), PNODE, and LaSDI. In this setting, adaptive sampling can improve both the learned autoencoder representation and the latent dynamics model.

2. We agree that adaptive sampling did not improve accuracy when the uniformly-trained autoencoder was frozen. However, when using the more accurate adaptively-trained frozen autoencoder, we saw that adaptive sampling reduced maximum prediction error from 58.9% to 37.7%, and 99th-percentile error from 42.9% to 23.0% (Table 7).

The uniformly-trained autoencoders have large peak reconstruction errors (27.9% on average), which creates two difficulties: (i) the inaccurate autoencoder limits prediction accuracy, and (ii) adaptive sampling may select parameters with latents that are difficult to model. When the autoencoder is frozen, adding these trajectories cannot improve the autoencoder representation and we cannot alter the latents so that they are easier for our dynamics models to capture.

We acknowledge the negative result with frozen, uniformly-trained autoencoders. However, our additional experiment demonstrates that our adaptive sampling criterion does improve the learned dynamics model with a frozen autoencoder, provided that the autoencoder is sufficiently accurate.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response

We thank the reviewers for their helpful feedback, and believe that addressing the reviewers' feedback has substantially strengthened the manuscript. In response to reviewer suggestions, we have revised the manuscript in two main ways:

Positioning. To address novelty concerns of Reviewer s9Yg, our introduction now explicitly acknowledges Takagi–Sugeno, LPV, and TPWL. We also describe dLDS and DVBF-LL, which were in our original related works section. Our contribution is the non-intrusive parametric ROM methodology for high dimensional ODEs. We introduce our method for joint training of autoencoders and continuous-time latent dynamics, while also providing analysis and adaptive training. We do not claim the affine mixture-of-experts model class itself is new.

Experiments. We have added additional baselines and several adaptive sampling ablations. Namely,

1. We added benchmarks from the original PNODE and GPLaSDI papers. We consider the 1D forced inviscid Burgers' equation from the original PNODE manuscript, and the viscous 2D Burgers' equation from the original GPLaSDI manuscript (Section 6.2, Table 1, Appendix B.2).

2. We have also applied our adaptive criterion to PNODE (Section 6.3, Table 2 and Appendix B Figure 8), run random-sampling CABLE ablations (Appendix B.3.1, Table 7), and run Reviewer s9Yg's suggested frozen-autoencoder ablations on Vlasov (Appendix B.3.1, Table 7).

In the below responses we discuss the main results and address reviewer-specific concerns. We provide the specific location(s) for each change. We have also uploaded a PDF with highlighted changes in purple. We note that we slightly shortened Section 6.1 and combined our original appendices B.1–B.4, which are not highlighted.

AuthorsOpenReview ↗

Follow-up on Revised Manuscript and Responses

We again thank the reviewers for their feedback.

As the discussion period is ending soon, we would greatly appreciate any further feedback from the reviewers on our revision. We are happy to address outstanding concerns, questions, or conduct small follow-up experiments (if feasible) before the deadline.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.