Gaussian Partial Information Decomposition: Bias Correction and Application to High-dimensional Data

Recent advances in neuroscientific experimental techniques have enabled us to simultaneously record the activity of thousands of neurons across multiple brain regions. This has led to a growing need for computational tools capable of analyzing how task-relevant information is represented and communicated between several brain regions. Partial information decompositions (PIDs) have emerged as one such tool, quantifying how much unique, redundant and synergistic information two or more brain regions carry about a task-relevant message. However, computing PIDs is computationally challenging in practice, and statistical issues such as the bias and variance of estimates remain largely unexplored. In this paper, we propose a new method for efficiently computing and estimating a PID definition on multivariate Gaussian distributions. We show empirically that our method satisfies an intuitive additivity property, and recovers the ground truth in a battery of canonical examples, even at high dimensionality. We also propose and evaluate, for the first time, a method to correct the bias in PID estimates at finite sample sizes. Finally, we demonstrate that our Gaussian PID effectively characterizes inter-areal interactions in the mouse brain, revealing higher redundancy between visual areas when a stimulus is behaviorally relevant.

Paper

References (54)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer YrjF8/10 · confidence 3/52023-07-04

Summary

The paper proposed a new method for partial information decomposition (PID) on multivariate Gaussian distributions. The issue of bias was discussed, and a correction method was provided. The method was tested on synthetic canonical examples and real data.

Strengths

1. The introduction clearly lays out the problem. 2. The paper is extremely well written with notations clearly defined. 3. The method extends prior work on PID with new properties. 4. The method is rigorously tested on simulated data. 5. The method solves an important problem, namely “the extent to which one region’s activity uniquely explains that of another, while excluding information corresponding to spontaneous behaviors” as stimulus could make two regions seem correlated.

Weaknesses

1. The paper does not provide enough real data to show effectiveness in real applications. 2. Testing on higher dimensional settings would be important.

Questions

1. Would non-modeled confounding factors affect the results? Summary after rebuttal The added simulations for testing higher dimensional data further justify the high score I gave.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The paper does not have negative societal impacts.

Reviewer DDWz6/10 · confidence 3/52023-07-05

Summary

Partial Information Decompositions (PIDs) play an essential role in neuroscience research. One constraint of the broader usability of PIDs is the computational difficulty of computing PIDs for high-dimensional neural recordings. To address this concern, the authors propose a method to compute and estimate a PID efficiently. More specifically, they restrict the optimization space of PIDs to jointly Gaussian, which reduces the number of optimization variables, allowing them to compute PIDs for much higher dimensionalities of neural data. Then, the authors show their method could be written out in closed form and solved by projected gradient descent, and they use nine examples with increasing complexity to show their method has the ability to recover ground truth and stability over increasing dimensionality. The authors also claim it's the first time to correct the bias and variance of estimates in a heuristic way. Finally, the authors evaluate the performance of their method on both synthetic and real neural data.

Strengths

* Computational scalability of ~PID with a basic property called additivity. * The authors' exposition of their method and experiments is clear. * The first time to raise the issue of bias in PID estimates.

Weaknesses

* No analysis is done to show the stability of $\delta$-PID over increasing dimensionality. According to the paper, there are two differences between $\delta$-PID and $\sim$-PID. The first one is the "additivity" property, which is clearly shown by Examples 8-9 in Section 4. The second one is that $\sim$-PID uses an exact upper bound, while $\delta$-PID uses an approximate upper bound. However, no analysis is performed to show how the second difference will affect the performance. In other words, people may be curious about the stability of $\delta$-PID over increasing dimensionality if we use examples with no "additivity" property ($\delta$-PID and $\sim$-PID both agress with the ground truth). Currently, it's hard to see the differences between $\delta$-PID and $\sim$-PID when applying to high dimensional neural data. I think this point may determine whether the proposed contribution is timely and impactful or a solid technical upgrade without practical consequences in neuroscience.

Questions

* Does $\delta$-PID also agree with the ground truth even when distributions are not Gaussian? * Could you please discuss how the "additivity" property would affect the analysis of communication among brain regions?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

I have highlighted technical limitations and weaknesses above. I have nothing further to add here.

Reviewer wULj6/10 · confidence 3/52023-07-07

Summary

The authors propose a new, efficient method for computing Partial Information Decompositions (PIDs) on multivariate Gaussians. They build their approach around the $\sim$-PID approach, as this allows them to preserve an additivity property (allowing PIDs to be computed on independent systems separately and then added later). They present a number of canonical examples for Gaussians, shaw that their Gaussian PID works even when distributions are non-Gaussian, and show an example of its use on real neural data from the Allen Institute. Finally, they address the issue of bias in PID estimates, propose a bias-correction method, and evaluate it empirically.

Strengths

The paper is very clearly written and the topics well explained. The literature review seems comprehensive, and although of somewhat specific topical interest, the work seems original and useful.

Weaknesses

Given one of the major motivating factors given in the introduction is the need for efficient estimators of PID that can accommodate higher dimensional neural data, I was disappointed that in the end the authors chose to apply their method to a PCA-reduced version of the Allen Institute neuropixel probe data... wasn't the point to be able to address higher dimensional problems? It's thus not completely clear to me that this method, as is, has delivered on the promise of providing a PID approach more applicable to data with thousands of neurons than other PID methods.

Questions

Can the authors method run on thousands of neurons, consistent with the original motivation they provide? What are the limitations on dimensionality more explicitly for this method? What is the computational complexity? Does the author's method require using PCA or having high firing rate neurons? How does it deal with low firing rate neurons or highly inhomogenous Poisson variables?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors adequately address limitations, although it would be good to hear more related to the questions above.

Reviewer Tp5P6/10 · confidence 4/52023-07-20

Summary

This article proposes an upper bound on mutual information (more precisely: on the "unique information" / "union information" that appear in "Partial Information Decomposition") that is easier to compute. This is done by replacing an infimum over all distributions matching given marginals by an infimum over only the ones that are Gaussian. Toy examples (with Gaussian distributions) are given, and an experiment on real data is performed. Further theory also includes considerations about biases arising from covariance estimators.

Strengths

The paper is mostly clearly written, with simple examples to help the reader follow. The paper is self-complete, with reminders of definitions. Considering covariance estimator biases is a plus.

Weaknesses

EDIT: after discussion with the authors, I see that I had misunderstood the scope and contributions of the paper. The weaknesses below do not longer hold (or not as strongly). ----------- The main issue is that the definition of the upper bound actually supposes that the distribution MXY that is studied is Gaussian, at least marginally. Indeed the infimum is performed over all possible distributions Q_MXY that are Gaussian and whose marginals satisfy Q_MX = P_MX and Q_MY = P_MY, which implies that P_MX and P_MY have to be Gaussian (otherwise the infimum is performed over an empty set). As a consequence, this upper bound definition cannot be applied to datasets that are not Gaussian (at least marginally). This crucial point is not discussed in the paper. Actually, from Sections 3.2 and 5, it seems that the upper bound is estimated based on covariance matrices only, without using distribution Gaussianity. If this is the case indeed, then the definition could be changed, to rely directly on these covariance matrices and removing the Gaussianity assumption. This however would probably require to significantly rewrite the paper. Another issue is the lack of sufficient validation, either theoretically or experimentally, in particular regarding distributions that are not Gaussian (i.e., really not Gaussian). This comment could have been tempered if there had been a discussion about general approximative Gaussianity of distributions in the field of study, but there is none. Also, the impact of the introduction of such a Gaussianity constraint in the upper bound estimation should be studied closely: how realistic is this assumption, how far are real optimal Q_MXY from their Gaussian versions, how tight is the upper bound thus obtained, etc. If this was a theoretical paper, I would expect theoretical results (guarantees such as a bound on the error made by the upper bound, or proving the conjectures, etc.). If this was an experimental paper, I would expect an extensive validation on many distributions for which the actual mutual informations are estimated (brute force or known solutions) in settings as varied as possible (not just normal distributions + a single real dataset). But the paper, in its current state, does meet these expectations.

Questions

I would be willing to revise my score if the points in the section above were significantly addressed. EDIT: they were.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Cf the validation issue in the weakness section above.

Reviewer jMyH7/10 · confidence 3/52023-07-25

Summary

First, I would like to thank the authors for the work they put into contributing to the field. In their paper, the authors provide a new method to estimating a well-known measure, the Partial Information Decomposition (PID). Within the topic of PIDs, their method computes a specific version of PID, the ~-PID. While methods to compute the ~-PID already exist, the authors' method is advantageous in that 1) it is more efficient to compute (quadratic in number of variables instead of exponential) and 2) in that it corrects for the bias. The author's then go on to show with simulations and examples that 1) their choice of PID method, the ~-PID, outperforms other PID methods 2) the bias correction works empirically 3) their method is stable even for increasing dimensionality 4) their method can be applied on data which is not Gaussian distributed such as simulated Poisson data and real neural recordings (neuropixel)

Strengths

- The paper addresses a topic which is becoming more and more relevant as the number of neurons that can be recorded from simultaneously is increasing: The question of how information is distributed among neurons and brain areas. - The paper is written in a very clear and structured way which successfully guides the reader through every logical step. - The mathematical formulation is sound - The quality of figures is excellent

Weaknesses

- The paper is an improvement of an already existing method which makes it important and publishable yet disqualifies it for a very high grade. - The claim of generalizability to non-Gaussian data is not backed up strongly enough empirically or mathematically (refer to "Questions") - Small mistakes/typos in lines: 177, 206, 322

Questions

- In the list of contributions, the authors state in line 50 that they are able to reduce the number of optimization variables from exponential to quadratic. Where can I see this in the equations and derivations? - The authors show that their method works on non-Gaussian data by applying it to simulated Poisson data and neuropixel spike count data. They explain this by stating that the Gaussian distributions is a good-enough approximation to the Poisson distribution. It is known, however, that neural firing does not follow a Poisson distribution. This is especially pronounced when recorded via 2photon imaging for which the distribution is strongly inflated at zero. Xue-Xin Wei et al. 2020 propose a zero-inflated gamma model to accurately capture calcium imaging traces. Is it feasible for the authors to test their method on such bimodal distributions which might be harder to approximate with a Gaussian distribution? If the rebuttal time for such experiments is too short, can the authors comment on the expected outcome?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

2 fair

Limitations

The authors discuss and point out most limitations. The claim that their method works on non-Gaussian data is only backed by an experiment with Poisson data, however. In my opinion, there is still room for the method to fail for other relevant distributions (refer to point "Questions") which would be worth mentioning as a (possible) limitation.

Reviewer Tp5P2023-08-11

Thank you for your reply. I had misunderstood the scope of the paper indeed. I thought the Gaussian case was trivial and that the point was to reduce non-Gaussian examples to Gaussian ones as a first approximation (through Definition 2). Now I understand that even the Gaussian case is complex. I thought that for Gaussian distributions, the minimizer Q in Definition 1 would necessarily be Gaussian, and thus introducing Definition 2 was relevant only for non-Gaussian distributions (hence the misunderstanding on the scope). If the conjecture is true, i.e. that minimizers Q in the Gaussian case are Gaussian indeed, then the two definitions are identical. If I understand well, the contribution of the paper in that case is not to bring a new definition, but to make explicit use of the knowledge that the minimizer is Gaussian, to help the optimization process, while this was not done in previous papers. Am I right?

Authorsrebuttal2023-08-13

Response to Comment by Reviewer Tp5P

We are very grateful to the reviewer for their response. > I had misunderstood the scope of the paper indeed. I thought the Gaussian case was trivial and that the point was to reduce non-Gaussian examples to Gaussian ones as a first approximation (through Definition 2). Now I understand that even the Gaussian case is complex. > > I thought that for Gaussian distributions, the minimizer Q in Definition 1 would necessarily be Gaussian, and thus introducing Definition 2 was relevant only for non-Gaussian distributions (hence the misunderstanding on the scope). The reviewer is absolutely correct that the Gaussian case is also non-trivial in general. They also correctly recognize that it is _not known_ whether the minimizer Q is guaranteed to be Gaussian if P_MXY is jointly Gaussian. Even for Gaussian P_MXY, this is only known in a few specific cases such as when M is scalar, or if P_MXY satisfies a condition shown in [12]. > If the conjecture is true, i.e. that minimizers Q in the Gaussian case are Gaussian indeed, then the two definitions are identical. If I understand well, the contribution of the paper in that case is not to bring a new definition, but to make explicit use of the knowledge that the minimizer is Gaussian, to help the optimization process, … Indeed, we do _not_ provide a new definition. Rather, the central contribution of our paper is to develop and evaluate an efficient method to estimate the PID for Gaussian P_MXY, by restricting the space of the minimizer Q to Gaussian variables. As the reviewer states, this would result in the correct solution if our conjecture is true. By assuming Q to be Gaussian, we can make use of the properties of Gaussian distributions (e.g., closed-form expressions for mutual information) to develop an efficient projected gradient descent optimizer, and then check how well we perform (please see _”Evaluating the Gaussianity assumption”_ below). > … while this was not done in previous papers. Am I right? A previous paper by Venkatesh and Schamberg [12] used a similar technique of restricting Q to be Gaussian for computing a different PID definition called the delta-PID. Our work makes significant advances over [12], on multiple fronts: 1. We compute a different PID definition, the ~-PID, which satisfies better properties than the delta-PID used in [12]. Most importantly, the ~-PID satisfies a very fundamental property called **additivity** (please see our response to reviewer DDWz on why additivity is fundamental). 2. We examine several examples (Examples 5-10), increasing in complexity and dimensionality, to show that restricting Q to be Gaussian is reasonable, and allows us to recover the ground truth. [12] only tested this for the simplest case of Gaussians with scalar M, and fails in more complex cases (Example 8; Fig 2 in the paper). 3. Our method is faster (please see Fig 7 in the PDF attached to the overall rebuttal). 4. Our method is capable of computing higher dimensionalities (as described in our response to reviewer DDWz). 5. Finally, we consider the problem of _estimation_, i.e., computing the PID from a covariance matrix that is estimated from real data, and show how bias in PID estimates can be corrected. This is not addressed in [12]. These points are summarized in the overall rebuttal near the top of this page. **Evaluating the Gaussianity assumption** Since we do not have a guarantee that the minimizer Q is actually Gaussian for all Gaussian P_MXY, we test what the reviewer asked in their original review, but for _Gaussian_ P_MXY. That is, we test: “how realistic is this assumption [of restricting Q to be Gaussian], how far are real optimal Q_MXY from their Gaussian versions”. We consider a number of examples of increasing complexity (Examples 5-10), where the ground truth PID values are known (starting with scalar M [13], and then using the additivity property). We show that our method is able to recover the ground truth, proving that for all the examples we consider, the optimal Q _is_ in fact Gaussian, and hence our conjecture still stands. Incidentally, our method is also applicable to non-Gaussian P_MXY, because it only relies on the covariance matrix, as noted by the reviewer in their original review. So we further tested whether it was reasonable to apply our method to Poisson distributions (now extended to more non-Gaussian distributions; see overall rebuttal). Here, we do not have ground truth, so we instead compared our results with another ~-PID estimator [20] (which works only for discrete distributions with limited support). We showed that our PID estimator with Q-restricted-to-be-Gaussian comes very close to the discrete estimator of [20] on a multivariate Poisson distribution P_MXY (Fig 5 in the paper). We then used this as a basis to apply our PID estimator to real neural data, and demonstrated its utility in providing new insights about interactions between brain regions (Fig 6 in the paper).

Reviewer Tp5P2023-08-16

Thank you for your detailed answer. Since complexity and dimensionality are relevant in the claims, I have questions regarding the optimization. 1. From the last figures you have provided, it looks like the computational complexity is cubic. In the paper, it seems that 2 matrix inverses (or pseudo-inverses) and one SVD need to be perform, which would explain this cubic complexity. Do you confirm? Where are the computational bottlenecks? It could be relevant to add a line in the paper stating that complexity. 2. How stable is the computation of the gradient? In particular, the projection (Equation 13 + Appendix lines ~470-500) replaces negative eigenvalues by 0. What about eigenvalues that are close to 0? Are there very small eigenvalues that could be on either side or the 0 threshold, up to statistical fluctuations (such as +-1 sample), and what is the impact of such variations (as they change the dimension of the projection subspace, and consequently the projected matrix as well)? Actually, have you checked the spectrum?

Authorsrebuttal2023-08-17

Response to Comment #2 by Reviewer Tp5P

> From the last figures you have provided, it looks like the computational complexity is cubic. In the paper, it seems that 2 matrix inverses (or pseudo-inverses) and one SVD need to be perform, which would explain this cubic complexity. Do you confirm? Where are the computational bottlenecks? It could be relevant to add a line in the paper stating that complexity. Yes, the computational complexity of each gradient descent iteration is cubic in the dimensionality of M, X and Y. Suppose for simplicity that M, X and Y all have the same dimensionality $d$. Then, the computational complexity of the objective, the gradient and the projection operator are all $\mathcal O(d^3)$. Since each of these steps has a similar complexity, we don’t believe that there is any particular bottleneck. We will certainly add a statement about computational complexity to the revised paper, and include the timing analysis from the attached PDF as well. > How stable is the computation of the gradient? In particular, the projection (Equation 13 + Appendix lines ~470-500) replaces negative eigenvalues by 0. What about eigenvalues that are close to 0? Are there very small eigenvalues that could be on either side or the 0 threshold, up to statistical fluctuations (such as +-1 sample), and what is the impact of such variations (as they change the dimension of the projection subspace, and consequently the projected matrix as well)? Actually, have you checked the spectrum? **Regularization for stable inverses:** The main source of instability in computing the gradient are the matrix inverses in Eq (55) in the Appendix. These are regularized by adding an identity matrix multiplied with 1e-7 (as mentioned in the Appendix, lines 491-492). We are not completely sure what the reviewer means regarding the eigenvalues, but we try to offer a response based on our best understanding: we first note that negative eigenvalues are set to zero on $\Sigma_{XY|M}$, whereas the optimization variable $\Sigma$ is an off-diagonal block of this matrix (please see Eq (56)). Setting several eigenvalues to zero in $\Sigma_{XY|M}$ will also affect the rank and the invertibility of the matrix $S = I - \Sigma\Sigma^T$, whose inverse appears in Eq (55). This is why we have the regularization in place: $S^{-1} = ((1 + \epsilon)I - \Sigma\Sigma^T)^{-1}$, with $\epsilon = 10^{-7}$. In the projection step, we only set negative eigenvalues to zero. We do not apply a threshold to zero-out any small positive eigenvalues. We have checked the spectrum of $\Sigma_{XY|M}$, and it is certainly possible for it to have several small positive eigenvalues. Along with the eigenvalues set to zero, this may affect the rank and invertibility of $S$, as mentioned above. However, we observed that the aforementioned regularization solves this problem in practice. We have now also added an extra regularization term to the projection operator for the matrix inverses in Eq (58). **Role of RProp:** We also note that RProp is designed to suppress the learning rate of any components along which the gradient is unstable and/or fluctuating a lot. Since fluctuations will likely change the sign of the respective gradient component rapidly, RProp will automatically suppress the learning rate $\eta$ along that component specifically (please see Eq (61)). **Not using SGD:** Please also note that this is _not_ a stochastic gradient descent method. The objective is fixed for a given covariance matrix (even if the covariance itself is estimated from data), and there is no concept of a mini-batch, so the stability of the gradient under stochasticity of mini-batches is not a concern (in case this was a point of confusion). **Ground truth comparison is the ultimate test:** Ultimately, the true test of whether our implementation works lies in whether it recovers ground truth. We have shown that our method recovers ground truth in various cases (Examples 5-10), including when M, X and Y each have a dimensionality of 1024 (please see the updated result in an official comment at the top of this page). We request the reviewer to let us know if we have answered their questions, and if they would kindly consider raising their score, based on a renewed understanding of the scope and contributions of our paper.

Reviewer Tp5P2023-08-17

> We request the reviewer to let us know if we have answered their questions, and if they would kindly consider raising their score, based on a renewed understanding of the scope and contributions of our paper. Yes, thank you for all your answers, I will raise my score. A last question, regarding RProp: the paper (and the comment above) mention RProp (which dates from 1992), but the paper referred to ([33]) is RMSProp (2012). Which one do you use? To me, they are very different techniques.

Authorsrebuttal2023-08-17

Response to Comment #3 by Reviewer Tp5P

We are extremely grateful to the reviewer for their time and effort in providing a helpful review, and for having a constructive discussion with us. We also thank the reviewer for raising their score. > A last question, regarding RProp: the paper (and the comment above) mention RProp (which dates from 1992), but the paper referred to ([33]) is RMSProp (2012). Which one do you use? To me, they are very different techniques. We thank the reviewer for raising this point. We do use RProp, not RMSProp, and we agree that they are different. We will use the correct reference to _Riedmiller and Braun (1993)_ shown below, to make this clear. Riedmiller, Martin, and Heinrich Braun. "A direct adaptive method for faster backpropagation learning: The RPROP algorithm." _IEEE international conference on neural networks_. IEEE, 1993. We had been unable to find this reference at the time of submission, so we had cited [33] instead, which points to a set of lecture notes and a webpage that include an explanation of RProp.

Reviewer jMyH2023-08-14

Clarifications and extra experiments

I thank the authors for their clarifications concerning the impact of the work and for their additional experiments showing the (limited) generalizability to non-Gaussian data. I have raised the score from 6 to 7.

Reviewer YrjF2023-08-15

Response to Rebuttal

The added simulations for testing higher dimensional data further justify the high score I gave.

Authorsrebuttal2023-08-15

Extending our method to a dimensionality of 1024

Dear reviewers, We were able to extend the dimensionality of our method to 1024 in Example 10 (results shown below). The dimensionality here refers to the number of dimensions in each of the vectors, M, X and Y. This is therefore equivalent to having a total of 3072 neurons across three brain regions, or 1024 PCA components in each brain region, with a covariance matrix of size 3072x3072. We believe this addresses Reviewer wULj's question of whether our method can achieve the stated goal of being able to analyze thousands of simultaneously recorded neurons. Although this is a simulated example and not real neural data, we believe this example is more informative about our method's capabilities since it has known ground truth and allows for computing error metrics (as shown below). This also expands the difference between the capabilities of our method and the delta-PID of [12] (as requested by Reviewer DDWz), which failed at a dimensionality of 256 on the same example, and ran much slower (our method takes roughly the same amount of time at 1024 dimensions, as theirs takes at 64 dimensions). The earlier error in our method at 512 (in the PDF attached to the overall rebuttal) stemmed from the projection step of our projected gradient descent, where we had missed adding a regularization term while inverting a matrix. This did not cause problems until a very high dimensionality, since the condition number of the matrix being inverted got worse with dimensionality. We have now added an adaptive regularization, which ensures that the projection step works correctly at any dimensionality. **Updated Results** The following table shows the results from Example 10 (Fig 3 in the paper), extended to 1024 dimensions in each of M, X and Y. The results are shown for successively doubling at a gain value of 3.0 in X_1 (i.e., for successively doubling the last data point in Fig 2(left) in the paper). The errors shown are taken with respect to ground truth, and represent the maximum error across all PID components. We have not yet tried running this analysis at a higher dimensionality of 2048, since we wanted to share this update as soon as possible. | Dimension | I(M;(X,Y)) | UI_X | UI_Y | RI | SI | Max absolute error (bits) | Max relative error | Compute time (s) | |-----------|-------------|------|------|----|----|----------------------|----------------|------------------| | 2 | 2.73 | 1.16 | 0.29 | 1.0 | 0.28 | 6.49e-08 | 2.35e-07 | 0.04 | | 4 | 5.46 | 2.32 | 0.58 | 2.0 | 0.55 | 1.26e-07 | 2.28e-07 | 0.04 | | 8 | 10.92 | 4.64 | 1.17 | 4.0 | 1.11 | 2.41e-07 | 2.18e-07 | 0.05 | | 16 | 21.84 | 9.29 | 2.34 | 8.0 | 2.21 | 4.97e-07 | 2.25e-07 | 0.06 | | 32 | 43.68 | 18.58 | 4.68 | 16.0 | 4.42 | 1.01e-06 | 2.29e-07 | 0.11 | | 64 | 87.35 | 37.15 | 9.36 | 32.0 | 8.84 | 2.05e-06 | 2.32e-07 | 0.29 | | 128 | 174.70 | 74.30 | 18.72 | 64.0 | 17.68 | 4.13e-06 | 2.34e-07 | 1.11 | | 256 | 349.40 | 148.60 | 37.44 | 128.0 | 35.36 | 8.28e-06 | 2.34e-07 | 3.59 | | 512 | 698.81 | 297.21 | 74.88 | 256.0 | 70.73 | 1.66e-05 | 2.35e-07 | 59.14 | | 1024 | 1397.61 | 594.41 | 149.75 | 512.0 | 141.45 | 3.32e-05 | 2.35e-07 | 473.42 | We request the reviewers to consider whether these results (in addition to the extra analyses in the overall rebuttal) make our paper worthy of a higher score.

Authorsrebuttal2023-08-18

Follow up comment to Reviewer DDWz

We request Reviewer DDWz to let us know if our responses have addressed their concerns, and to kindly consider whether our paper is now worthy of a better score. We also request the reviewer to get back to us soon, since author responses will be closed after August 21st, 1pm EDT. In addition to the overall rebuttal and the reviewer-specific rebuttal, we would also like to draw the reviewer's attention to the analysis posted in an official comment at the top of this page, which extends our simulation in Example 10 to 1024 dimensions in each of M, X and Y. This further shows how our method is more capable and applicable than the $\delta$-PID of [12], which fails at a dimensionality of 256 in the same experiment.

Authorsrebuttal2023-08-18

Follow up comment to Reviewer wULj

We request Reviewer wULj to let us know if our responses have addressed their concerns, and to kindly consider whether our paper is now worthy of a better score. We also request the reviewer to get back to us soon, since we author responses will be closed after August 21st, 1pm EDT. In addition to the overall rebuttal and the reviewer-specific rebuttal, we would also like to draw the reviewer's attention to the analysis posted in an official comment at the top of this page, which extends our simulation in Example 10 to 1024 dimensions in each of M, X and Y (equivalent to 3072 total neurons). We believe this addresses the reviewer's central point about extending our method to thousands of neurons, as explained in the official comment.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC