acceptodds
Under review as a conference paper at ICLR 2027

Scalable Bayesian Time Warping at Single-Neuron Resolution

Abstract

Neuroscience analyses often require trial-averaging to determine reliable neural signals. However, trial-to-trial heterogeneity in the timing of neural dynamics often limits the efficacy of trial-averaging. In many cases, this heterogeneity may be linked to latent events, such as decision time, which lack obvious behavioral correlates. Here, we formulate a scalable Bayesian approach to time warping to resolve this challenge. We leverage newly developed methods for efficient Gaussian process regression, introduce a prior over monotonic piecewise-linear functions, and develop a parallel, blocked Hamiltonian Monte Carlo-within-Gibbs sampling approach to learn a full posterior distribution over model parameters. Moreover, we show this approach is capable of identifying warping function posteriors for individual neurons. On synthetic data, this capability can be used to identify latent subpopulation structure invisible to existing methods. On data from rat orbitofrontal cortex during a temporal wagering task, population-level fits favor more complex warps than most individual neurons require, providing evidence for heterogeneous single-neuron alignment surrounding a latent decision. Additionally, fitting to individual neurons sharpens firing patterns obscured by a shared population warp. In all, our results illustrate that Bayesian time warping is a practical tool for neural data analysis that not only mitigates trial-to-trial heterogeneity, but can leverage it to reveal previously hidden structure in single-neuron dynamics.

Primary area:
applications to neuroscience & cognitive science
Keywords:
neural population dynamics, temporal alignment, Gaussian processes, MCMC, probabilistic models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_gatxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a Bayesian method that aligns neural spike trains across trials. The key idea is to assume that each neuron has a template describing its typical response over time, which is warped differently across trials. The warping function is shared across all neurons, but the authors also fit each neuron separately to allow different alignments. The template function has a Gaussian process (GP) prior followed by a softplus transformation, while the warping function is assumed to be monotonic and piecewise linear, with segment lengths generated by applying a softmax transformation to Gaussian variables.

The posterior distributions of the template and warping functions are inferred using Hamiltonian Monte Carlo (HMC) within Gibbs sampling. The conditional independence structure allows parallel sampling, and the authors also use parallel tempering to explore different possible solutions. Finally, the warping complexity is selected using Pareto-smoothed importance sampling leave-one-out cross-validation (PSIS-LOO).

On synthetic data, the authors introduce a measure called "warp swap dissimilarity" to cluster neurons with similar warping functions. On rat orbitofrontal cortex (OFC) data, they find that most individual neurons prefer simple shift warps, while the full population prefers more complex piecewise-linear warps.

Critical strengths

This paper addresses an important problem in neural data analysis. Differences in neural response timing may contain useful information about the underlying neural circuits. Estimating these differences can help align neural responses across trials and improve our understanding of neural dynamics.

The proposed method successfully recovers the underlying template and warping functions on synthetic data, as shown in Figure 2.

The proposed warp swap dissimilarity measure works well for identifying groups of neurons with similar warping functions on synthetic data (Figure 3). It achieves better clustering performance than directly comparing the posterior means of the warping function

Critical weaknesses

In the rat decision-making experiment, the authors use PSIS-LOO to estimate model performance when leaving out one trial at a time. However, the warping function is trial-specific. When a whole trial is left out, the remaining trials provide no direct information about its warping function. It is therefore unclear how the method predicts the neural activity on the held-out trial. It is possible that the prior distribution of warping function would carry cross trial information that can be used by the method, but that does not seem to be clear to me.

In the evaluation on the rat decision-making data, the authors state that they use the trial-averaged neural activity from even trials as a template and generate predictions for odd trials. However, it is unclear how the warping functions for the odd trials are obtained. Do the authors first fit the model using all trials, including the odd trials, to estimate their warping functions before evaluating the model?

The comparison between different models and fitting methods in Figure 3E is interesting. In particular, the improvement in clustering performance when using the Bayesian method instead of a point estimate is promising. However, I have a few concerns: 1)The comparison with Williams et al. (2020) needs more details. What template functions were used in the baseline model? Was regularization included, and how were the regularization parameters selected? 2)The posterior-mean warping function already achieves good clustering performance. Although the proposed "warp swap dissimilarity" further improves performance, it is unclear how important this improvement is in practice. It would be interesting to test more challenging scenarios where the posterior-mean method performs poorly but using the full posterior distribution provides a clear advantage. This would help demonstrate when modeling posterior uncertainty is necessary.

The model includes neuron-specific template functions and trial-specific warping functions, which introduce many free parameters. When applying such a flexible model to neural recordings, an important question is whether the available data are sufficient to reliably estimate these parameters. It would be helpful to investigate how the estimation accuracy depends on the number of recorded neurons and trials, as well as the number of spikes per neuron. For example, the authors could perform simulations with different numbers of neurons and trials to study parameter recovery and uncertainty. This analysis would also help us understand the bias-variance tradeoff of the model and identify the conditions under which the proposed method works reliably.

Questions for authors

1.How is PSIS-LOO used for single-neuron model selection in Figure 4? In particular, how are the trial-specific warping functions handled when a whole trial is held out?

2.In the evaluation shown in Figure 4, how are the neural activities from odd trials used during model fitting? Are the odd trials used to estimate their own warping functions before evaluating predictive performance?

  1. Could the authors provide a more challenging example to demonstrate the advantage of using the full posterior distribution of warping functions over their posterior means, beyond the results in Figure 3E?

  2. How much neural recording data is needed to reliably estimate the model parameters? In particular, how does estimation accuracy depend on the number of recorded neurons, trials, and spikes per neuron?

5 comments
AuthorsOpenReview ↗

Response to Reviewer gatx [1 / 2]

Thank you for your thoughtful review of our work.

How is PSIS-LOO used for single-neuron model selection in Figure 4?

We apologize for an oversight in the original draft of the paper, which naively applied the same PSIS-LOO model selection procedure for the population model and the single-neuron model. As you correctly point out, leaving out (neuron, trial)-tuples is an ill-defined procedure in the single-neuron case because the model is unable to predict data on entirely unseen trials given only data from a single neuron. Therefore, to perform model selection at the single-neuron level, we instead leave out (timebin, trial)-tuples. This LOO quantity is slightly optimistic due to non-negligible autocorrelation between adjacent timebins, but it allows us to perform model selection at the single-neuron level. With this revised LOO procedure, there was a shift in the optimal complexity histogram towards more complex warps (Figure 4C). However, the majority of neurons still preferred a shift (31/48 neurons) and none of the chosen example neurons had their optimal complexity change under the new procedure. Therefore, the qualitative conclusions of Section 3.3 still hold. The new complexity distribution also makes the trial-split log-likelihood result more significant (new p-value of 0.00977). The controls did not have their optimal complexity histograms change with the new procedure (Figure S6). We have revised the text to include the single-neuron PSIS-LOO procedure, and we have explicitly noted the limitation regarding the autocorrelation of timebins.

In the evaluation shown in Figure 4, how are the neural activities from odd trials used during model fitting?

We compute a split-trial Poisson log-likelihood as a metric to compare the raw data, the population fits, and the single-neuron fits. The procedure is as follows: we first fit the model to all trials, take the aligned trial average of the even trials as a pseudo-template, pass this pseudo-template through the warps learned for the odd trials, and compute the Poisson log-likelihood of the odd trial data under that model. As your questions suggest, this is not a held-out measure of predictive performance, as we use the odd trial data to learn the necessary warping functions. Rather, it quantifies how well the aligned (or raw) trial average explains single-trial activity. In other words, we are computing how similar the activity looks on single trials compared to the trial-average. We view this as an extension of the approach used in Figure 2F-H of Williams et al. (2020) to Poisson data along with extra robustness conferred by using the trial average of the even trials to explain the observed activity on odd trials. We have added further clarification of the approach to the manuscript on lines 478-480 and 1157-1159.

The comparison with Williams et al. (2020) needs more details.

The cross-validation procedure for hyperparameter tuning introduced in Williams et al. (2020) is not feasible for per-neuron fitting, as it relies on holding out entire trials. In order to select reasonable hyperparameters for the affine warping model, we performed 3-fold cross-validation when fitting to the entire population and fix these hyperparameters for all single-neuron fits. We saw qualitative examples where the affine warping model was able to identify warping functions close to the data-generating warping function when fit to single neurons (Figure S5C). We also found examples where the affine warping model did not identify warping functions close to the data-generating warping function (Figure S5B). Notably, in these cases, the posterior distribution identified by Bayesian time warping often provided reasonable coverage of both the data-generating warping function as well as the MLE identified by affine warping (Figure S5B). Critically, affine warping could still support clustering if its errors were correlated within a subpopulation which shared a warping function. However, the quantitative (Figure 3E) and qualitative data (Figure S5A-B) suggested that its errors were not correlated within a subpopulation. We have expanded Appendix B.2 (lines 1105-1110) to provide more details regarding the affine warping fitting procedure and Section 3.2 (lines 410-415) to clarify the shortcomings of affine warping for clustering.

AuthorsOpenReview ↗

Response to Reviewer gatx [2 / 2]

Could the authors provide a more challenging example to demonstrate the advantage of using the full posterior distribution of warping functions over their posterior means, beyond the results in Figure 3E?

This is a great question. If the posterior is unimodal, we envision that the posterior distribution can likely be well-summarized by the mean of the posterior samples. However, if the posterior exhibits multimodality, the mean can be a poor summarization of the posterior. As an example, Figure S3 demonstrates the substantial multimodality observed in the warping functions when the model is fit to the OFC data. In this setting, the posterior mean of the warping functions is very low probability under the posterior and resembles none of the drawn samples, which would be an incorrect characterization of the warping functions relevant to that trial. We revised Figure S3 to include a visualization of the posterior mean compared to the samples from the posterior to highlight this effect. This means that we expect clustering on the posterior means to be appropriate in cases with well-behaved, unimodal posteriors (which is why it can achieve a max ARI of 1.0 on some of the synthetic subpopulation datasets), but will struggle in multimodal posteriors. The degree of warping multimodality present in all experimental datasets is an empirical question out-of-scope for this work, but we present the warp swap dissimilarity as a possible clustering approach which is designed to be robust to multimodality by directly using samples from the posterior. Outside of clustering, we highlight that the full posterior allows for us to propagate alignment uncertainty into downstream analyses via variance-tagged rasters (Figure 3A, middle) or posterior-smoothed spike counts (Figure 3A, bottom). Future work will assess if incorporating this uncertainty provides advantages for downstream tasks (such as encoding/decoding analyses) relative to only using the posterior mean.

In particular, how does estimation accuracy depend on the number of recorded neurons, trials, and spikes per neuron?

We completely agree that characterizing the bias-variance tradeoff and reliability of flexible statistical models is very relevant for practical applicability to neuroscience. To our knowledge, we are the first to rigorously tackle this question for time warping models. We highlight Section 3.1 and Figure 2B-C, which emphasize coverage as the relevant quantity of interest when fitting a Bayesian model, rather than simply accurate estimation of the ground truth parameters. We show that, even when shrinking the number of trials by 10x, the template estimates still exhibit ideal coverage behavior by widening the posterior (Figure 2B-C, left). Similarly, even when shrinking the number of neurons by 10x, the warping estimates also exhibit ideal coverage behavior by widening the posterior (Figure 2B-C, right). Tracking the dependence on signal-to-noise ratio is a great suggestion also brought up by reviewer yM9r. Rather than firing rate, a more appropriate measure of signal-to-noise ratio for our model is given by the output scale of the Gaussian process. Intuitively, this quantity determines the dynamic range of a neuron's template - at lower output scales, there is less variance in the neuron's firing rate, which means there is less ability to distinguish different timepoints in the template. This naturally makes the templates and warping functions less identifiable. Qualitatively, we observe appropriate widening of the posterior and increased prior influence as the output scale decreases in synthetic data. We include a new figure tracking the inferred posterior over template and warping functions for an example neuron/trial as we vary the output scale of a synthetic dataset (Figure S4).

I appreciate the authors’ clarifications on how they perform cross-validation on single-trial level and how the odd-trial data is used. However, I feel the answers to the bias-variance question are not sufficient. Could the author vary systematically the number of neurons/trials included to see how their method based on full posterior distribution outperform the methods that are only based on posterior means?

Given the fact that it is probably not possible to use this method to predict how a new trial is warped. I wonder whether this method can shed any new light on the mechanism underlying the neuroscience process. For example, in the OFC data the author analyzed, when using this method to identify individual neurons’ warping, is it possible to use this heterogeneity to gain more insights on what are the task-relevant signals and how this brain region perform this task?

Reviewer_yM9rOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors present a Bayesian approach towards time-warping that is able to discover the warping functions of individual neurons. Overall, I think it is a nice methods development paper, but have some concerns about its practical utility for neuroscientists.

Critical strengths

-Generally clear storyline

-Methodological tour-de-force

Critical weaknesses

-As the authors acknowledge when discussing limitations, the template function for each neuron is conserved across all trials (assuming all trials are identical). This seems to really limit the scope of applicability, as the vast majority of experiments have multiple conditions in which we would expect different template functions.

-Continuing along the thread of practical use for neuroscientists, it seems like author’s primary argument for a scientific use case is that it could enable clustering neurons based on warp timing, rather than based on the underlying signals that neurons relate to. But it’s unclear to me that if two neurons had similar timing, but based on completely different underlying signals (template functions in this case), how this would be suggestive of this being a meaningful ‘cluster’ in a circuit (e.g. what would give rise, or be the meaning, of this cluster)?

Questions for authors

-The regime in which this method will be accurate is currently unclear - e.g. in simulations, what are the firing rates you are using? If you go to a regime with very low firing rates, so that the time warping is inaccurate, does your Bayesian approach allow a user to know there is very high uncertainty?

-This is more minor, but I had some confusion when reading about different model variants. For instance, I don’t think it was very clear what the “population” fit was in the real data figure? (Though I think from context that it’s where each neuron gets its own template, but the warping function is shared). Similarly, the simulation figures seem to allude that having more neurons is helpful for fitting (when making the transition from using a simulation where there are multiple neurons to doing individual-neuron fits), but why is this helpful if each neuron gets its own template?

5 comments
AuthorsOpenReview ↗

Response to Reviewer yM9r

Thank you for your very kind words about our work.

The template function for each neuron is conserved across all trials (assuming all trials are identical). This seems to really limit the scope of applicability, as the vast majority of experiments have multiple conditions in which we would expect different template functions.

We agree that there are many experiments with variable conditions across trials which would result in distinct templates. However, it is still common for experimenters to record multiple trials per condition to improve the signal-to-noise ratio via trial-averaging. In these cases, it is straightforward to apply our method separately to each subset of trials with similar conditions, which makes the shared template assumption better specified. In these cases, there will be fewer trials to constrain the model, but our model is able to represent this with wider template posteriors (Figure 2B-C).

 It’s unclear to me that if two neurons had similar timing, but based on completely different underlying signals (template functions in this case), how this would be suggestive of this being a meaningful ‘cluster’ in a circuit (e.g. what would give rise, or be the meaning, of this cluster)?

Our thesis is that standard analysis approaches for identifying functional subpopulations of neurons rely on similarity in response shape and timing between units (e.g. Gaussian Process Factor Analysis or Tensor Component Analysis). In higher-order areas, it has been observed in prior experimental studies that subgroups of neurons may be tuned to different events, and neurons in the same subgroup may display heterogeneous tuning to the same events (Hirokawa et al., 2019; Posani et al., 2026; Osako et al., 2026). Therefore, we believe that a method based solely on similarity in response timing is a novel contribution that may be able to identify groups of neurons responsive to the same latent event, even if members of that subpopulation exhibit extremely diverse tuning profiles to the underlying event. As a concrete example, neurons that ramp up to decision and neurons that ramp down to decision can look different to shape-based clustering, but would share a warping function locked to latent decision time. We now introduce this argument on lines 369-379 to better motivate Section 3.2.

 If you go to a regime with very low firing rates, so that the time warping is inaccurate, does your Bayesian approach allow a user to know there is very high uncertainty?

The average firing rate for a neuron in the simulations for Figure 2A range from 0.255 spikes per bin to 2.405 spikes per bin. Tracking the dependence on signal-to-noise ratio is a great suggestion, also brought up by reviewer gatx. Rather than firing rate, a more appropriate measure of signal-to-noise ratio for our model is given by the output scale of the Gaussian process. Intuitively, this quantity determines the dynamic range of a neuron's template - at lower output scales, there is less variance in the neuron's firing rate, which means there is less ability to distinguish different timepoints in the template. This naturally makes the templates and warping functions less identifiable. Qualitatively, we observe appropriate widening of the posterior and increased prior influence as the output scale decreases in synthetic data. We include a new figure tracking the inferred posterior over template and warping functions for an example neuron/trial as we vary the output scale of a synthetic dataset (Figure S4). A related question we investigate in Section 3.1 and Figure 2B-C is if the "Bayesianness" of the model correctly inflates uncertainty as the evidence per-parameter shrinks by subsampling the number of trials and number of neurons by 10x. We illustrate that the model correctly inflates its posterior intervals to obtain near-optimal coverage behavior as the evidence per-parameter shrinks.

I had some confusion when reading about different model variants.

We apologize for the confusion about the model variants. The population fit in Figure 4 refers to the model when sharing warping functions across all neurons, in contrast to fitting the model to each neuron independently (which allows for neuron-specific warping functions). In Section 3.1 and Figure 2B-C, subsampling neurons shrinks the evidence per warping parameter when we assume that all neurons share the same warping functions. On the other hand, subsampling trials shrinks the evidence per template parameter because we assume that all trials share the same template functions. We show that subsampling does not compromise the coverage behavior of our method (Figure 2B), but subsampling trials leads to wider posterior intervals around the template functions (Figure 2C, left) and subsampling neurons leads to wider posterior intervals around the warping functions (Figure 2C, right). We revise the text on lines 473-474 to make the model variants clearer.

Thank you for response.

As a followup, the regime in which the model will be accurate remains unclear - the listed spikes per bin in your simulation would correspond to a range of 12.75 to 120.25 spks/sec using the 20ms bins that you use in real data, which I'm guessing is likely about 10 times the actual firing rates.
It's a very nice method - it would just be helpful to have some intuition of the scenarios in which the model will be accurate and we can glean scientific insights from it. While it is of course comforting that the model will produce a wide posterior when there is limited data, users should at least have some idea of when it's likely to have not-super-wide posteriors and scientific insight can be more easily gleaned. This relates to firing rates, and also to trial counts (which would be limited if you need to divide trials by conditions, as mentioned in your first response)

AuthorsOpenReview ↗

Response to Comment by Reviewer yM9r

It's a very nice method - it would just be helpful to have some intuition of the scenarios in which the model will be accurate and we can glean scientific insights from it.

Thank you for the follow-up and continued enthusiasm about our method.

We agree that it is helpful to provide intuition to users for how the model's estimates change with various data conditions. We explore three data conditions in simulated data: subsampling trials, subsampling neurons, and varying the output scale. The key intuition for the model's behavior in these conditions comes from the fact that the two parameter types (template functions and warping functions) are constrained by different sources of evidence.

The template functions are shared across trials, so their identifiability is related to the spikes per neuron, accumulated across trials. Therefore, subsampling trials by 10x (from 40 trials to 4 trials) reduces the evidence per template parameter by 10x. We show in Figure 2B (left) that the model exhibits near-optimal coverage behavior in this scenario by widening its posterior estimates. Encouragingly, the posterior over template functions after subsampling trials is still informative (Figure 2C, left).

The warping functions are shared across neurons, so their identifiability is related to the spikes per trial, accumulated across neurons. In this case, subsampling neurons by 10x (from 20 neurons to 2 neurons) reduces the evidence per warping parameter by 10x. We show in Figure 2B (right) that the model also exhibits near-optimal coverage behavior in this scenario by widening its posterior estimates. Again, the posterior over warping functions after subsampling neurons is still informative (Figure 2C, right).

To your point about firing rates, a 10x reduction in firing rate reduces the number of spikes per neuron and spikes per trial by 10x, approximately the same as the reduction in spike count achieved by subsampling. Our subsampled conditions therefore approximate the full dataset at 1.3-12 Hz, and the posteriors remain informative there. However, your original point about the signal-to-noise dependence of the method was astute and well-taken, inspiring us to explore the dependence on the output scale. This quantity reflects the temporal modulation of the templates' firing rate. Intuitively, as the output scale decreases, the template functions get flatter - this means that distinguishing timepoints in the template becomes more difficult and makes both the template functions and warping functions less identifiable. At an output scale of 0.1 (Figure S4A, first column), the template firing rate of an example neuron is only modulated from 0.65-0.75 spikes/bin (mean of 0.7 spikes/bin, or 35 Hz using 20ms bins). Critically, even though this neuron is firing at 35 Hz, there is not enough temporal modulation to meaningfully distinguish different timepoints of the template, leading to uninformative posteriors for the template and warping functions (which are also heavily regularized towards the prior). We hope that this clarifies why firing rate alone is not a sufficient metric to quantify the identifiability of our model. As we get to an output scale of 10 (Figure S4A, third column), the template firing rate is modulated from 0.4-1.2 spikes/bin (mean of 0.8 spikes/bin, or 40 Hz using 20ms bins). By this point, there is enough temporal modulation to result in meaningful-but-wide posteriors. Our simulated data was generated with output scales ranging from 100 to 1000 (Figure S4A, fourth and fifth columns), meaning that the model is still able to provide meaningful estimates despite a 10x reduction in the temporal modulation of the template.

To our knowledge, we believe our simulation results to be the first attempt to systematically characterize the reliability of statistical time warping models for neuroscience. In practice, posterior width depends on several interacting factors (e.g. firing rate, temporal modulation, trial count, and population size) that will vary across datasets. Our real data results suggest that modest amounts of data are sufficient when templates are strongly modulated: we show that you can fit the model to single neurons independently and get an informative estimate of the posterior (Figure 4D, rows 5-6). If you believe it would strengthen the paper, we would be happy to include a condensed version of these arguments in a "Practical Considerations" section of the Discussion, summarizing how these factors affect identifiability.

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

In this work, the authors proposed a scalable Bayesian-based time-warping framework. The proposed method includes Gaussian Process w/ firing-rate templates, and parallel Hamiltonian Monte Carlo-within-Gibbs sampling to infer trial alignments that fit to individual neurons. Experiments on synthetic dataset and real-world neural data on rat orbitofrontal cortex recordings manifest that the inferred single-neuron alignment posteriors uncovers unique temporal heterogeneity and firing patterns thatn the baseline methods with no single-neuron resolution.

Critical strengths
  1. The problem this work is focused on: preventing the heterogeneity information lost from trial-averaging in experimental data analysis, plays an important role in neuroscience research.
  2. The proposed method is first-principle grounded and has a efficient sampling solution.
  3. The experiemnts on synthetic and real-world rat orbitofrontal cortex datasets are solid.
Critical weaknesses
  1. Could there be more experimental results across more subjects and sessions? currently there has only one real-world cortex dataset. More dataset results can potentially strengthen the scientific conclusions in the work.
  2. Are there more experimental comparisons with baseline methods only have shared population-level warping resolution?
Questions for authors

I have no more questions here. Please refer to my weaknesses section to see my other questions or concerns.

1 comment
AuthorsOpenReview ↗

Response to Reviewer Apnf

Thank you for your positive comments about our work.

Currently there has only one real-world cortex dataset. More dataset results can potentially strengthen the scientific conclusions in the work.

We agree that applications to further datasets are a necessary next step to make neuroscientific conclusions. Our primary goal in this work was to develop and validate the method, which requires access to ground truth. Additionally, we wanted to introduce the idea of heterogeneous warping functions within a recorded population. Therefore, we focused on synthetic experiments evaluating coverage, robustness, and clustering accuracy (Sections 3.1-3.2). We include the OFC dataset as a proof-of-concept that the method recovers structure in real population data and a result showcasing heterogeneous single-neuron warps compared to a population fit.

Are there more experimental comparisons with baseline methods only have shared population-level warping resolution?

We compare to affine warping (Williams et al., 2020) which optimizes template and warping functions via maximum likelihood estimation in Section 3.2 and Figure 3E. The key result is that, when fit to single-neurons that share the same ground-truth warp, affine warping does not learn similar warping functions. This means that it cannot be reliably used for subpopulation identification/clustering. However, we agree that additional comparisons on the real data would strengthen the paper. We now include real data comparisons between our method and affine warping fit to the whole population in Figure S8. Affine warping, like the population fit in Figure 4D, finds qualitatively different structure than that identified by the single-neuron fits in Figure 4D. We also note that we quantitatively compare the Bayesian model fit to single-neurons to the Bayesian model fit to the whole population as well (Figure S7).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.