acceptodds
Under review as a conference paper at ICLR 2027

On Feature Learning in Flow Matching under Different Couplings

Abstract

Flow matching learns velocity fields from paired noise and data samples without simulating trajectories during training. We discuss feature learning of flow matching, focusing on the difference between the independent and optimal-transport (OT) couplings for the same source and target distributions. To understand the difference, we employ a theoretical model, where the data lie on an unknown low-dimensional subspace, with a non-Gaussian distribution along one unknown direction (feature) and Gaussian variation in the orthogonal directions. We show that, near the target endpoint, the nonlinear component remains nonzero under OT but vanishes under independent coupling, while both couplings have the same velocity component normal to the subspace. To examine how this difference changes feature learning, we analyze joint learning of the support and feature direction at a fixed flow time using a model with one cubic neuron. Although nonlinear feature learning becomes increasingly slow near the target endpoint under independent coupling, OT avoids this slowdown, presenting more eminent feature learning. We also empirically confirm this advantage of OT in joint training across flow times, using narrow windows near the endpoint or separate features for each time bin. These results suggest that OT coupling may benefit the learning of nonlinear structure in the target density near the endpoint.

Primary area:
generative models
Keywords:
flow matching, feature learning, multi-index models, low-dimensional subspace, optimal transport conditional flow matching, theory of deep learning, toy models of neural networks

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_f4nQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how independent and optimal transport (OT) couplings affect feature learning in flow matching. The authors consider a simplified data model supported on a low-dimensional subspace with a nonlinear feature along one unknown direction. They show that near the data endpoint, the nonlinear component of the population velocity vanishes under independent coupling but remains nonzero under OT coupling. They further analyze gradient flow training with a cubic-neuron model and show that feature learning becomes increasingly slow under independent coupling. The theoretical findings are supported by synthetic experiments and controlled experiments using MNIST and CIFAR-10 decoder manifolds.

Critical strengths
  1. Interesting theoretical insight. The paper provides a new perspective on the benefits of OT coupling beyond the commonly discussed reduction in training variance and straighter transport paths. In particular, the analysis connects the choice of coupling to the strength of nonlinear feature-learning signals and the resulting training dynamics.

  2. Rigorous theoretical analysis. The paper establishes quantitative differences between independent and OT couplings, including the near-endpoint scaling of nonlinear signal strength and feature-learning time. The synthetic experiments provide supporting evidence for these theoretical predictions.

Critical weaknesses
  1. Restrictive assumptions in the theoretical analysis. The main theoretical results rely on strong simplifying assumptions, including a low-dimensional linear subspace, a single nonlinear feature direction, and a highly simplified learning model. In particular, Theorem 2 considers a single cubic neuron trained at a fixed flow time under population gradient flow. While these assumptions make the analysis tractable, it remains unclear to what extent the conclusions generalize to practical flow matching models with more complex data distributions and neural network architectures.

  2. Limited connection between the theoretical findings and practical performance. The paper does not establish a clear theoretical or empirical connection between its feature-learning results and the performance of practical flow matching models, which typically use shared parameters and multi-time training. Although the authors introduce theoretical measures of nonlinear signal strength and feature-learning time, it remains unclear whether these quantities predict meaningful improvements in training efficiency or generation quality. Ideally, the analysis would identify a theoretically interpretable and empirically measurable quantity that correlates with practical performance, and validate this relationship in standard flow matching settings. Without such evidence, the practical significance of the theoretical findings remains limited.

Questions for authors

Please see the weakness

3 comments

Thank you for the helpful comments. We appreciate your recognition of the analysis linking coupling-dependent nonlinear signals to feature-learning times.

W1 (Assumptions)

it remains unclear to what extent the conclusions generalize to practical flow matching models with more complex data distributions and neural network architectures.

Our theoretical setting assumes data supported on a low-dimensional linear subspace, but related low-dimensional assumptions are often used in feature-learning theory [1, 2]. For example, multi-index models assume that the output depends on only a few linear projections of the input, while single-index models restrict this dependence to one direction. In our setting, this subspace model allows us to separate support recovery from nonlinear feature learning.

Moreover, Figures 3–4 extend the comparison to models with features shared across times and CNNs on curved decoder manifolds, and the new Figure 5 (Appendix C.4) uses a standard two-layer tanh MLP with a non-polynomial link. These experiments provide evidence beyond the fixed-time cubic-neuron setting, although extending the theoretical guarantees to these models requires further analysis.

[1] Damian et al. Neural Networks can Learn Representations with Gradient Descent. COLT, 2022.
[2] Jimmy Ba et al. High-dimensional Asymptotics of Feature Learning: How One Gradient Step Improves the Representation. NeurIPS, 2022.

Theorem 2 considers a single one-cubic neuron

A single cubic neuron provides a minimal model for analyzing the relationship between nonlinear signal strength and feature-learning time. For the symmetric experimental target, the residual after removing the linear component is odd, making degree three the lowest nonlinear degree that can correlate with it. Symmetry alone does not guarantee nonzero cubic correlation; Theorem 2 explicitly assumes conditions that ensure it near the endpoint. We have added this explanation to the main text.

We distinguish the signal comparison from the training-time guarantee. Under its distributional assumptions, Theorem 1 compares nonlinear signals in the population velocity without restricting the learner's architecture. Theorem 2 gives a training-time guarantee for a cubic neuron. The data-generating function HH need not be cubic, provided it satisfies the theorem's assumptions, but training times for more general networks require a separate analysis.

W2(a) (Multi-time training)

practical flow matching models, which typically use shared parameters and multi-time training.

We agree that the role of feature sharing needs clarification. Figure 3(b) in Section 4 compares full-window and endpoint-window training on the same target using a cubic-neuron model. Figure 3(c) compares models with separate feature directions for different time bins and a shared support basis. Together, these results suggest that, when the feature direction is shared across times, signals at earlier times can mitigate slow feature learning near the endpoint.

W2(b) (Practical performance)

it remains unclear whether these quantities predict meaningful improvements in training efficiency or generation quality.

The quantity our analysis identifies is the nonlinear signal energy EC(t)\mathcal E_C(t) (Eq. (12)): the mean squared part of the population velocity that no linear predictor can fit. It is defined for any target and architecture, and can be estimated from the gap between the best linear and the best nonlinear regression loss at time t. For the cubic model, the relevant part of this signal is the residual cubic correlation κC(t)\kappa_C(t) (Eq. (20)). In our controlled experiments, the coupling with the larger signal learns the feature faster: for the cubic-neuron model (Figures 2–3) and, in the new Figure 5, for a standard two-layer tanh MLP. We have not tested practical models or generation quality, and we do not claim such gains (Section 5).

Thank you for the detailed clarification and additional experiments. However, my main concern remains the practical significance of the findings. While the analysis establishes a connection between nonlinear signal strength and feature-learning time, it is still unclear whether this translates into meaningful improvements in training efficiency or generation quality for practical flow matching models. I believe an experiment on a more realistic benchmark with a standard flow matching architecture would substantially strengthen the paper. I will maintain my score.

Thank you for the follow-up. We respond to the remaining concern below.

Practical benefit of OT coupling. That OT coupling helps standard flow matching models is already documented in other papers: minibatch OT improves the efficiency of training and inference [1] and gives better low-cost sample generation on downsampled ImageNet [2]. These works attribute the gains to lower gradient variance and straighter flows. While our paper does not claim a new practical gain, it identifies an additional mechanism behind the known ones: the nonlinear signal near the data endpoint persists under OT and vanishes under independent coupling.

Why we use controlled experiments. Comparing the two couplings on a realistic benchmark would reproduce the results of [1, 2], but it could not separate our mechanism from variance reduction. Separating them requires the population velocity, which is available in our synthetic setting (the population-label control in Appendix C.4) but not on real data.

What the analysis predicts for practical models. In our experiments, the gap is small when the feature is shared over the full time window (ratio 1.21 in Figure 3) and large when training is concentrated near the endpoint or features are time-specific (ratios 46.4 and 3.65). The signal energy EC(t)\mathcal E_C(t) can be estimated on real data to test this prediction. Such a test at realistic scale is a natural next step, which we could not complete in this revision.

[1] Tong et al. Improving and Generalizing Flow-Based Generative Models with Minibatch Optimal Transport. TMLR, 2024. [2] Pooladian et al. Multisample Flow Matching: Straightening Flows with Minibatch Couplings. ICML, 2023.

Reviewer_7AmWOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This work studies the choice of coupling affects feature learning in FM. They build a framework where the source and target are fixed. OT and ind coupling have the same normal velocity but different non-linear signals in the subspace. When removing the best linear predictor, a non linear signal remains positive near the endpoint under OT. A one cubic model is analyzed with a joint support and feature learning under fixed time population gradient flow resulting in hitting time bounds of Θ(r2)\Theta(r^2) and Θ(r2ϵt−2)\Theta(r^2\epsilon_t^{-2}) for OT and ind couplings. Finally, validation is done with synthetic experiments with fixed and shared flow times and vector field learning on MNIST and CIFAR-10 VAE decoder manifolds.

Critical strengths
  • I like the distinction between learning where the data lie and learning how they are distributed within that space. It gives a useful explanation of how changing the coupling can change the learning problem, even though the target is the same.
  • the joint training analysis is another strength. The authors do not assume that the support is already known or that the linear weights are instantly fitted. They analyze these updates together with the nonlinear feature direction, which makes the result more informative than a comparison of the optimal velocity fields alone.
Critical weaknesses
  1. I find the theoretical separation interesting, but I'm not yet convinced the paper shows when the endpoint slowdown matters beyond the fixed-time setting. Theorem 2 considers population gradient flow with a specific cubic-neuron model and exact population OT. In the shared-feature experiment over the full time window, the two couplings learn on similar timescales. The difference becomes more apparent when training is restricted towards the endpoint or when different time bins have separate features. This leaves me wondering how much the slowdown matters once features can be learned at earlier times and reused later. The feature direction is also identifiable from the target covariance, so this is a result about how the specified dynamics use available information, rather than a general difficulty in recovering the direction. These are reasonable choices for a theoretical study, but the broader significance is still not clear to me

  2. I find the connection between the image experiments and the theory less convincing. Appndix C.3 uses the perturbation z0↦z0+0.35tanh⁡(z0)z_0\mapsto z_0+0.35\tanh(z_0). Since E[Stanh⁡(S)]>0\mathbb E[S\tanh(S)]>0, this introduces a first-order covariance change, which Assumption 3.1 explicitly removes. Testing beyond the assumptions is fine, but it makes the interpretation less straightforward. The density direction metric also does not subtract the best linear predictor, unlike the nonlinear signal studied in Theorem 1. Without a covariance controlled comparison, it is hard to tell how much of the observed advantage reflects the proposed nonlinear learning mechanism rather than changes already visible in second-order statistics.

  3. Related to 2, the errors are measured against method specific learned reference fields, but the reported evaluation does not establish how accurate these references are in the density direction. I would be more convinced by the comparison if their uncertainty were quantified relative to the errors being reported.

  4. Minor: fig 3 says both axis are log but density direction have linear vertical axis

Questions for authors
  1. how much do the two independently trained reference networks disagree in the density direction, compared with the prediction errors reported in fig 3? Could you report this disagreement together with the absolute reference field energies at the smallest noise scales?
  2. what numerical criterion is used to decide whether the warped latent coordinate is locally identifiable? What fraction of the held out samples passes this filter for each method and noise level, and are the same target examples retained across methods?
3 comments

Thank you for the careful review and specific comments. We appreciate your recognition of the distinction between recovering the support and learning the distribution within it.

W1(a) (Feature sharing)

how much the slowdown matters once features can be learned at earlier times and reused later.

With full-window sharing the gap is small (ratio 1.21 in Figure 3(b)). It grows when training concentrates near the endpoint (ratio about 46.4 for the narrowest window in Figure 3(b)) or when features are not shared across times (ratio about 3.65 for K=8 in Figure 3(c)). The endpoint slowdown therefore matters when time sampling is endpoint-focused or when the network uses time-specific features.

Previous work also reports more efficient training with OT than with independent coupling when parameters are shared across times [1]. Our analysis suggests a possible explanation that under OT, updates throughout the time interval, including near the endpoint, can contribute to learning a shared feature direction.

[1] Alexander Tong et al. Improving and Generalizing Flow-Based Generative Models with Minibatch Optimal Transport. TMLR, 2024.

W1(b) (Covariance)

The feature direction is also identifiable from the target covariance

Thank you for pointing this out. We agree that the feature direction can be identified from the covariance of the original target. To remove this information, we conducted an additional experiment with a target whose covariance is isotropic, as described below.

W2(a) (Covariance control)

this introduces a first-order covariance change.

We agree that the original perturbation introduces a first-order covariance change and therefore does not satisfy the orthogonality condition in Assumption 3.1. To address this concern, we added a separate synthetic experiment with identity target covariance (Appendix C.4, Figure 5).

Specifically, we replace tanh⁡(s)\tanh(s) by H(s)=tanh⁡(s)−csH(s)=\tanh(s)-cs, where c=E[Stanh⁡(S)]c=\mathbb E[S\tanh(S)], so that E[SH(S)]=0\mathbb E[SH(S)]=0. We then normalize the transformed coordinate to unit variance, giving Cov⁡(X1)=I8\operatorname{Cov}(X_1)=I_8. We also evaluate nonlinear residuals after separately removing the model's fitted affine component and the population velocity's best affine predictor. The population velocity is computed numerically, without a learned reference.

In addition, to separate signal strength from label noise, we also trained the independent model on numerically computed population velocities. Its final nonlinear residual NMSE drops from 11.31 to 0.431 but remains far above that of OT (1.2×10−31.2\times10^{-3}), indicating that label noise contributes to the gap but does not fully explain it (Appendix C.4).

With this target-covariance control, the nonlinear signal still decreases toward the endpoint under independent coupling but persists under OT, and the MLP learns the feature direction faster under OT. These observations are consistent with our nonlinear-signal analysis.

This additional experiment uses synthetic data. We did not rerun the decoder-manifold experiments with a covariance-controlled warp. We now explicitly describe those experiments as qualitative evidence, since they fall outside Assumption 3.1 and their density-direction errors include linear components.

W2(b) (Nonlinear metric)

The density direction metric also does not subtract the best linear predictor

We agree that the original density-direction error includes linear components and does not isolate nonlinear learning. The new Figure 5(c) in Appendix C.4 evaluates nonlinear residuals after separately removing the affine components of the model and population velocity fields. The learning gap remains under this evaluation.

W3 / Q1 (Reference stability)

how much do the two independently trained reference networks disagree in the density direction, compared with the prediction errors reported in fig 3?

At ϵ=10−3\epsilon=10^{-3}, the smallest CIFAR-10 noise scale in Figure 4 (formerly Figure 3), the squared disagreement between the two references is 0.75–0.80 of the model's squared error for the two FM methods and 2.3 for diffusion. All quantities below are measured along the density direction. Model error and reference energy are computed using the mean of the two references.

MethodSquared reference disagreementModel squared errorReference energyDisagreement / error
Independent FM1.03×10−31.03\times10^{-3}1.38×10−31.38\times10^{-3}4.61×10−34.61\times10^{-3}0.75
OT-CFM3.74×10−33.74\times10^{-3}4.68×10−34.68\times10^{-3}1.93×10−11.93\times10^{-1}0.80
Diffusion1.44×1031.44\times10^36.24×1026.24\times10^24.26×1034.26\times10^32.31

Absolute scales differ across methods (Appendix C.3, Eq. (164)). Model error is measured against the mean of the two references. If their errors are independent, the mean has error variance D/4, where D is the squared disagreement, i.e. 19%, 20% and 58% of the model error for the three methods; shared bias is not captured. The lower relative error of OT-CFM in Figure 4 comes from a reference energy about 42 times larger along the density direction (disagreement is 2% of that energy), not from a smaller absolute error, which is 3.4 times larger than for independent FM.

W4 (Axes)

fig 3 says both axis are log but density direction have linear vertical axis

Thank you for pointing this out. We have corrected the caption.

Q2 (Identifiability)

what numerical criterion is used to decide whether the warped latent coordinate is locally identifiable?

We consider the warped latent coordinate locally identifiable if the squared norm of its unit vector projected onto the retained right singular subspace of the decoder Jacobian is at least 0.9.

What fraction of the held out samples passes this filter for each method and noise level, and are the same target examples retained across methods?

All 256 held-out samples (100%) pass the filter for every method and every plotted noise scale in Figure 4, on both MNIST and CIFAR-10. For each seed, the target examples and identifiability mask are shared across methods. We clarified these points in Appendix C.3.

Dear Reviewer 7AmW,

Thank you again for your detailed review. As the discussion period ends soon, we would like to check whether our responses address your concerns. In brief:

  • Covariance control and nonlinear metric (W2): a new synthetic experiment with identity target covariance, evaluated on nonlinear residuals against numerically computed population velocities, with a population-label control that removes label noise (Figure 5, Appendix C.4).
  • Reference reliability and identifiability (W3, Q1, Q2): the disagreement table, the criterion, and the retained fraction are now in Appendix C.3, and the decoder experiments are described as qualitative evidence.
  • Relevance beyond fixed time (W1): Section 4 now quantifies when the slowdown matters (alignment-time ratios 1.21, 46.4 and 3.65).

We are happy to clarify any remaining points.

Best regards,

Authors

Reviewer_fwL4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper discusses feature learning of flow matching and focuses on the difference between independent and OT coupling of the source and target distributions. The paper adapts the single-index framework from feature learning to flow matching and compare the speed of learning the support and feature direction under the two couplings. The main difference are as follows: The nonlinear component remains nonzero under OT but decreases under independent coupling for polynomial nonlinearity. The velocity labels have zero conditional variance under OT but have nonzero variance along the feature direction under independent coupling. In both theoretically and empirically, independent coupling learns endpoint features more slowly than OT coupling. Even though both couplings recover the support quickly, support recovery alone does not imply the efficiency of learning the nonlinear structure within it.

Critical strengths
  1. The paper is clearly organized, the theoretical comparison between the two classical types of couplings is directly provided in theorems, giving quantitive understandings to readers. The comparison is comprehensive, covering the nonlinear signal, the label noise, the joint support and the speed of feature learning.

  2. The perspective of comparing the two classical couplings in a specific feature learning framework beyond the common convergence rate analysis is novel and interesting.

Critical weaknesses
  1. The presentation of the paper can be further improved. I would suggest extract the numerical experiment Section 3.5 out of Section 3 and become a separate new section. The sketch of derivations are a bit too long, I believe the sketch can be made more concise with keeping the key ingredients. It would be beneficial to add a brief introduction of the setup (like the dataset and the experimental framework) of the numerical experiments at the beginning of Section 3.5 for better understanding.

  2. It would be more informative to add a toy experiment or a simple case study to visualize the differences discussed in the feature learning results, including the support and the feature direction.

Questions for authors
  1. In Figure 2, why are the curves of independent coupling not monotonic? What is the implications of the turning pints of these curves? Is this a common phenomenon for independent coupling?

  2. Why do we consider the cubic neuron, not neurons of other (higher or lower) orders in Section 3.4? Could the results in Section 3.4 generalize to other polynomial neuron?

  3. Is it feasible to add a toy example to visualize the differences of the two couplings in the feature learning results? Can the theoretical results be verified on larger datasets like ImageNet?

3 comments

Thank you for the helpful comments and suggestions. We appreciate your recognition of our feature-learning perspective and comparison of the two couplings.

W1 (Presentation)

I believe the sketch can be made more concise with keeping the key ingredients.

Thank you for your suggestion. We made the sketch concise for readability. In addition, we have added numbered outlines at the beginning of the proofs of Theorems 1 and 2, summarizing the main steps and pointing to the relevant lemmas and equations. We believe these revisions improved the readability of the proofs.

extract the numerical experiment Section 3.5 out of Section 3 and become a separate new section

It would be beneficial to add a brief introduction of the setup (like the dataset and the experimental framework)

Following your suggestion, we have moved the numerical experiments from the former Section 3.5 to a separate Section 4. The section now begins with a brief introduction to the data and experimental setups.

W2 / Q3(a) (Visualization)

a toy experiment or a simple case study to visualize the differences discussed in the feature learning results

Is it feasible to add a toy example to visualize the differences of the two couplings in the feature learning results?

Thank you for the suggestion. We have added Figure 1 at the beginning of Section 4 to visualize learning under OT and independent coupling, which is inspired by Figure 2 of Ba et al. [1]. Under both couplings, model samples quickly approach the V-shaped target support, but only OT captures the concentration around the target mode within the training period shown. Alongside the sample trajectories, a separate panel compares the target and final model densities. Appendix C.5 provides the setup details. This example illustrates how rapid support recovery can coexist with slower density learning, and how the coupling changes the latter.

Q1 (Nonmonotonicity)

In Figure 2, why are the curves of independent coupling not monotonic?

Under independent coupling, the cubic signal in this experiment vanishes at both endpoints and is stronger at intermediate times (Appendix A.5, Eq. (73)). In Figure 3(a), formerly Figure 2(a), narrowing the training window first excludes weak early-time signals, increasing the RMS. Further narrowing excludes the stronger intermediate-time signals, decreasing it.

Let κ‾ind(a)\overline\kappa_{\mathrm{ind}}(a) denote the RMS signal under the uniform distribution on the window (a,1)(a,1). Then

=\frac{\overline\kappa_{\mathrm{ind}}(a)^2-\kappa_{\mathrm{ind}}(a)^2}{1-a}.$$ Thus, at a turning point, the squared signal at the window boundary equals its average over the window. This explains the nonmonotonicity in panel (a). The turning points of the learning-time curves in panels (b,c) also depend on the joint dynamics of the shared support and output weights. **Q2 (Cubic neuron)** > Why do we consider the cubic neuron, not neurons of other (higher or lower) orders in Section 3.4? For the symmetric target $H=\operatorname{He}_3$, the velocity residual after removing the best linear predictor is odd and therefore has zero correlation with quadratic features. **Degree three is the lowest nonlinear degree that can have nonzero correlation with this residual (Appendix A.5).** This argument relies on target symmetry. Theorem 2 also permits noncubic data-generating functions $H$ that satisfy its assumptions, but training times for activations of other degrees require a separate analysis. **Q3(b) (ImageNet)** > Can the theoretical results be verified on larger datasets like ImageNet? Computational constraints prevent us from conducting ImageNet-scale experiments for this revision. Our main contribution is theoretical, and the controlled experiments in Section 4 are designed to test its mechanism. [1] Jimmy Ba et al. High-dimensional Asymptotics of Feature Learning: How One Gradient Step Improves the Representation. NeurIPS, 2022.

Thanks for your response and it helps me to understand the paper better, especially the non-monotonicity part. I would like to keep my score.

Thank you for the reply and for the time you spent on our paper. We are glad that the explanation of the non-monotonicity was helpful. Your suggestions on the presentation, including the separate experiments section and the visualization in Figure 1, have made the paper clearer, and we appreciate them.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Reviewers and AC,

Thank you for your time and constructive feedback. We are encouraged by the positive comments on the feature-learning perspective on couplings (fwL4, f4nQ), the quantitative comparison of nonlinear signal strength and feature-learning time (f4nQ), and the joint analysis of support and feature learning (7AmW).

Here we respond to a common question raised by Reviewers 7AmW (W1) and f4nQ (W1, W2), and then summarize the revision.

Q: When does the endpoint slowdown matter beyond the fixed-time cubic-neuron setting?

A: The signal comparison in Theorem 1 does not depend on the learner: near the endpoint, the nonlinear part of the population velocity stays of constant order under OT and decays as εt2\varepsilon_t^2 under independent coupling. Theorem 2 shows, for a cubic neuron at a fixed time, how this difference changes the feature-learning time. Two sets of experiments indicate when it matters beyond this setting.

  • Sharing across flow times (Section 4, Figure 3). With a feature shared over the full time window, the gap is small (independent/OT alignment-time ratio 1.21). The ratio grows to 46.4 when training is restricted to the window (0.9,1)(0.9,1), and to 3.65 when the feature is not shared across eight time bins. The slowdown therefore matters when time sampling is endpoint-focused or when features are time-specific.
  • Beyond the cubic neuron (Figure 5, Appendix C.4). On a target with identity covariance, a standard two-layer tanh MLP learns the feature faster under OT. Training the independent model on population velocities removes label noise and reduces the gap, but does not close it.

Summary of the revision (changes are marked in red)

  • The experiments are now in a separate Section 4, which begins with the setups. The new Figure 1 visualizes support recovery and density learning (fwL4).
  • A new covariance-controlled experiment with an MLP, including a population-label control (Figure 5, Appendix C.4) (7AmW, f4nQ).
  • Appendix C.3 now gives the identifiability criterion, the retained fraction, and the disagreement between the reference fields. The decoder experiments are described as qualitative evidence (7AmW).
  • Section 3.4 explains the choice of the cubic neuron (all reviewers).
  • The derivation sketches are shorter, and the proofs of Theorems 1 and 2 begin with outlines (fwL4).
  • Former Figures 1-3 are now Figures 2-4.

We are happy to discuss any remaining questions during the discussion period.

Best regards, Authors of Submission 23123

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.