arxiv: 2605.03061 · v2 · submitted 2026-05-04 · 📊 stat.ML · cs.LG· q-bio.QM· stat.ME

Recognition: unknown

Dynamic Vine Copulas: Detecting and Quantifying Time-Varying Higher-Order Interactions

Houman Safaai , Alessandro Marin Vargas

Authors on Pith no claims yet

Pith reviewed 2026-05-08 17:10 UTC · model grok-4.3

classification 📊 stat.ML cs.LGq-bio.QMstat.ME

keywords dynamic vine copulastime-varying dependencehigher-order interactionsvine copulasconditional dependenceneuropixelscopula modelsmultivariate time series

0 comments

The pith

Dynamic Vine Copulas isolate time-varying higher-order conditional dependence by contrasting full vines against their pairwise-truncated versions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Dynamic Vine Copulas to model and diagnose how multivariate dependence structures evolve over time, including aspects like tail dependence and conditional interactions that standard correlation-based methods overlook. It fixes a vine structure and tracks changes in the pair-copulas, using a predictive score that measures the extra contribution from higher trees in the vine. This matters for systems like neural recordings where simultaneous activity across areas may involve conditional dependencies that change with behavior. In benchmarks, it correctly identifies shifts in copula families or parameters that Gaussian models miss, and on real data it finds a signal of higher-order interactions that is consistent across data splits.

Core claim

Dynamic Vine Copulas (DVC) apply vine copula constructions to time series by letting pair-copula parameters follow smooth trajectories or regularized family switches, and the key diagnostic is the difference in predictive performance between a full vine and a 1-truncated vine that retains only the first tree; under the simplifying assumption this difference recovers the higher-tree part of the vine total-correlation decomposition and serves as evidence for time-varying conditional dependence.

What carries the argument

The higher-tree diagnostic score obtained by held-out comparison of the full dynamic vine against its matched 1-truncated version, which quantifies the predictive contribution of conditional pair-copulas beyond the first tree.

Load-bearing premise

The approach requires choosing and fixing one vine factorization in advance and assumes that higher-tree conditional copulas are functions only of the conditioning variables through their univariate marginals.

What would settle it

Observing that the higher-tree score remains significantly positive even after decorrelating the variables or fails to generalize across held-out splits on the Neuropixels recordings would falsify the claim that it isolates genuine conditional dependence.

Figures

Figures reproduced from arXiv: 2605.03061 by Alessandro Marin Vargas, Houman Safaai.

**Figure 1.** Figure 1: DVC overview. (A) A 4-variable C-vine with root order (1, 3, 2, 4) decomposes the copula into first-tree pairwise edges, second-tree conditional edges, and a third-tree conditional edge. (B) DVC fixes one vine factorization across time and lets edge states evolve within it. DVC-smooth assigns each edge one family and a smooth parameter path; DVC-switch selects a temporally regularized family/parameter path… view at source ↗

**Figure 2.** Figure 2: Tail and family-switch dynamics. (A) Piecewise Student-t degrees-of-freedom at fixed correlation: analytic truth, empirical tail summary, and fitted tail-dependence trajectories. (B) Held-out copula NLL (lower is better) for temporal variants on the same tail sequence. (C) Clayton-to-Gumbel switch at fixed Kendall’s τ : DVC-switch represents the family change as one temporally regularized state path, while… view at source ↗

**Figure 3.** Figure 3: Agent interaction episodes. (A) Ground-truth recurrent schedule. (B) DVC-switch held-out score decomposition, with matched Win. vine and DVC-smooth controls shown only where explicitly labeled. (C) NLL gaps against Gaussian, 1-truncated, Win. vine, and smooth-DVC comparators. (D–E) Detection and order assignment over time windows: the main advantage is not binary detection, but separating pairwise from hi… view at source ↗

**Figure 4.** Figure 4: Four-phase detection showcase (d = 10, T = 60), showing DVC-switch together with the independent Win. vine control. Mean trajectories over available seeds; shaded bands show seed-to-seed standard deviation when multiple seeds are available. DVC and Win. vine curves are held-out score estimates; dashed oracle curves are population information values. This figure emphasizes relative detection and order decom… view at source ↗

**Figure 5.** Figure 5: Allen Visual Behavior Neuropixels (VBN) joint-DVC temporal validation. Dots denote session-seed fits from five random held-out splits of the same presentation-order windows. (A) Experiment time course: windowed mean post-stimulus activity for familiar and novel sessions (session-wise z score), with image-change and rewarded-event fractions on the right axis. (B) Joint-DVC temporal NLL gain and higher-tree … view at source ↗

**Figure 6.** Figure 6: Scenario×baseline summary. (A) Mean held-out NLL gap for the reported estimator in each scenario (DVCsmooth or DVC-switch for temporal settings; explicitly labeled windowed full-vine controls only for static/structural diagnostics) against representative comparators, including the independent Win. vine wherever a joint DVC is available. Colorbar clipped at 2.0, multiplicative-triplet cell annotated separa… view at source ↗

**Figure 7.** Figure 7: Higher-order structure and structural recovery. (A–C) Multiplicative triplet shown as conditional density slices: the unconditional Y –Z view is nearly uncorrelated, while splitting on the sign of X reveals strong positive and negative dependence (∆NLL = +0.316 nats over the Gaussian copula; full-vs-1-truncated gain = +0.179 nats). (D) Hub switching: independent windowed and regularized windowed vines both… view at source ↗

**Figure 8.** Figure 8: Dalgleish stimulation-aligned latent-state analysis. (A) Session-level baseline comparison for the selected non-targeted latent variant: the full vine has positive gaps relative to Graphical Lasso, Gaussian SSM, Gaussian copula, and the 1-truncated vine. (B) The gain decomposes into a near-zero pairwise-flexible component and an estimated higher-tree contribution that is positive on average, supporting but… view at source ↗

**Figure 9.** Figure 9: Allen Visual Behavior Neuropixels (VBN) cohort validation. (A) Session-level mean higher-tree held-out gap ∆b HO = NLL1-trunc − NLLfull for paired familiar and novel sessions from 8 mice. Bars show mean ± SD, dots are sessions, and gray lines connect paired mice. Values are session-summed nats across held-out presentations/windows, unlike the per-held-out-presentation values in view at source ↗

read the original abstract

Time-varying dependence is often modeled with dynamic correlations or Gaussian graphical models, but multivariate systems can change through tail behavior, asymmetry, or conditional structure even when correlations are nearly stable. We introduce Dynamic Vine Copulas (DVC), a temporal vine-copula framework for estimating and diagnosing sequence-wide non-Gaussian dependence. DVC fixes a chosen vine factorization for comparability; the framework applies to C-, D-, and R-vines, and our experiments use fixed-root-order C-vines. Pair-copula states evolve through smooth parameter trajectories or temporally regularized family-switching paths. The main diagnostic is a held-out comparison between a full vine and its matched 1-truncated version, which separates flexible first-tree pairwise dependence from evidence contributed by higher-tree conditional terms. At the population level, under a correct fixed vine and the simplifying assumption, this contrast equals the higher-tree component of a vine total-correlation decomposition; in finite samples, it is a predictive diagnostic. In controlled benchmarks, DVC detects Student-t degrees-of-freedom changes, Clayton-to-Gumbel switches, and recurrent conditional-interaction episodes missed or conflated by Gaussian dynamic baselines. The higher-tree score remains near zero in pairwise-only regimes and rises during conditional-interaction regimes. On Allen Visual Behavior Neuropixels data, DVC identifies a reproducible time-indexed higher-tree signal that is positive across held-out splits and vanishes under a decorrelated null, indicating simultaneous cross-area dependence. DVC therefore provides a flexible temporal copula model and an interpretable test of whether temporal dependence changes are pairwise or conditional.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a vine-based way to flag time-varying conditional dependence via a held-out full-versus-truncated likelihood gap, with decent synthetic checks, but the neural-data interpretation rests on an untested simplifying assumption.

read the letter

The main takeaway is that Dynamic Vine Copulas let you fit time-evolving pair-copulas inside a fixed vine and then use the held-out predictive gap to the 1-truncated version as a diagnostic for whether higher trees are adding real signal. This separates pairwise from conditional contributions in a way that standard dynamic correlation models do not directly provide. The controlled benchmarks on Student-t df changes, Clayton-Gumbel switches, and recurrent conditional episodes show the method can pick up dependence shifts that Gaussian baselines miss or conflate, and the null test on the Allen data is a reasonable sanity check. That part is useful and cleanly motivated from vine total-correlation ideas under the usual assumptions. The soft spot is exactly the one the stress-test flags: the gap is only guaranteed to isolate higher-tree conditional structure under the simplifying assumption and a correct vine, yet the paper shows no residual checks or non-simplified comparisons on the neural marginals or first-tree fits. If those are off, the positive score on Neuropixels could be driven by misspecification rather than genuine simultaneous cross-area dependence. The fixed vine choice and lack of sensitivity analysis add another layer of dependence on modeling decisions. This is for people who already work with copulas or multivariate time-series dependence and want a diagnostic tool beyond pairwise models, especially in neuroscience or similar domains. A reader who cares about interpretable higher-order interactions will get concrete ideas from the benchmarks and the held-out contrast. It deserves a serious referee because the core construction is grounded and the empirical tests are falsifiable, even though the application section would need more assumption validation to stand on its own.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces Dynamic Vine Copulas (DVC), a temporal vine-copula framework that fixes a vine factorization (e.g., C-vines) and lets pair-copula parameters or families evolve via smooth trajectories or regularized switching. The central diagnostic is the held-out log-likelihood gap between a full vine and its matched 1-truncated version, which is claimed to isolate higher-tree conditional contributions; under a correct fixed vine and the simplifying assumption this gap equals the higher-tree term in the vine total-correlation decomposition. Controlled benchmarks show detection of Student-t df changes, Clayton-Gumbel switches, and recurrent conditional episodes, while the Allen Visual Behavior Neuropixels application reports a reproducible positive higher-tree signal across held-out splits that vanishes under a decorrelated null, interpreted as evidence of simultaneous cross-area dependence.

Significance. If the central diagnostic is robust, DVC supplies a flexible, non-Gaussian alternative to dynamic-correlation or Gaussian-graphical models that can distinguish pairwise from conditional dependence changes over time. Strengths include the predictive (held-out) nature of the contrast, explicit null testing, and controlled benchmarks that isolate specific dependence regimes missed by Gaussian baselines. The framework's applicability across C-, D-, and R-vines and its parameter-free population-level link to vine total correlation (under the stated assumptions) are also positive features.

major comments (2)

[Abstract / application] Abstract and application section: the claim that the held-out gap 'indicates simultaneous cross-area dependence' on the Allen Neuropixels data rests on the simplifying assumption that higher-tree conditional copulas depend on the conditioning variables only through their marginals. No diagnostic (e.g., comparison to non-simplified vines, residual dependence tests, or sensitivity to marginal specification) is provided for the spike-count or LFP marginals, so the gap could be driven by first-tree misspecification rather than genuine conditional structure.
[Abstract] The finite-sample predictive diagnostic is presented as approximately equal to the higher-tree total-correlation term only 'under a correct fixed vine and the simplifying assumption.' Because the vine factorization is chosen by the user and the assumption is not validated, the quantitative link between the reported score and 'higher-order interactions' is weaker than stated for the neural-data regime.

minor comments (2)

The manuscript should clarify the exact temporal regularization strength and vine-root ordering choices used in the Neuropixels experiments, as these are listed among the free parameters.
Figure captions and table legends would benefit from explicit statements of the number of held-out splits and the precise null-construction procedure (e.g., how decorrelation is performed while preserving marginals).

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment point by point below, agreeing where the concerns are valid and outlining the revisions we will implement.

read point-by-point responses

Referee: [Abstract / application] Abstract and application section: the claim that the held-out gap 'indicates simultaneous cross-area dependence' on the Allen Neuropixels data rests on the simplifying assumption that higher-tree conditional copulas depend on the conditioning variables only through their marginals. No diagnostic (e.g., comparison to non-simplified vines, residual dependence tests, or sensitivity to marginal specification) is provided for the spike-count or LFP marginals, so the gap could be driven by first-tree misspecification rather than genuine conditional structure.

Authors: We agree that the interpretation of the higher-tree signal as evidence of simultaneous cross-area dependence relies on the simplifying assumption, which the manuscript states explicitly but does not validate with additional diagnostics for the neural marginals. We acknowledge that this leaves open the possibility of first-tree misspecification contributing to the observed gap. In the revision we will add a dedicated discussion of this limitation, including sensitivity analyses to alternative marginal specifications (e.g., different count models for spike data) and residual dependence checks where computationally feasible. We will also revise the abstract and application section to use more qualified language, replacing 'indicating' with 'consistent with' or 'suggestive of' simultaneous cross-area dependence. The decorrelated null, which preserves the empirical marginals while destroying dependence, provides partial protection against purely marginal-driven artifacts, as any such artifact would appear equally in the null distribution; we will emphasize this point in the revised text. revision: yes
Referee: [Abstract] The finite-sample predictive diagnostic is presented as approximately equal to the higher-tree total-correlation term only 'under a correct fixed vine and the simplifying assumption.' Because the vine factorization is chosen by the user and the assumption is not validated, the quantitative link between the reported score and 'higher-order interactions' is weaker than stated for the neural-data regime.

Authors: We concur that the population-level equality to the higher-tree component of the vine total-correlation decomposition holds only under a correctly specified vine structure and the simplifying assumption, and that the vine factorization is user-chosen. The manuscript already qualifies the held-out gap as a finite-sample predictive diagnostic rather than an exact decomposition. To address the concern directly, we will revise the abstract to foreground these caveats, stating that the diagnostic isolates higher-tree contributions under the stated assumptions and that the vine structure is fixed by the analyst. We will also expand the discussion section to clarify the distinction between the predictive utility of the contrast (which does not require the exact decomposition) and its interpretation as a quantitative measure of higher-order interactions, thereby tempering the strength of the claim for the neural-data application. revision: yes

Circularity Check

0 steps flagged

No significant circularity; held-out diagnostic is independent of in-sample fit.

full rationale

The paper's primary diagnostic is an explicit held-out predictive contrast (full vine vs. 1-truncated vine) on unseen data splits, which does not reduce to the in-sample parameter estimates by construction. The population-level link to the vine total-correlation decomposition is stated only under the standard simplifying assumption and a fixed vine structure, with the finite-sample result presented separately as a predictive diagnostic. No self-definitional equations, fitted parameters renamed as predictions, load-bearing self-citations, or ansatz smuggling appear in the derivation chain. The framework remains self-contained against external benchmarks and null models.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on the vine copula decomposition and the simplifying assumption; free parameters include the fixed vine structure choice and regularization strength for temporal trajectories, while no new entities are postulated.

free parameters (2)

vine factorization choice
Fixed chosen vine (C-vine with fixed root order) for cross-time comparability; selection affects which conditional terms appear in higher trees.
temporal regularization strength
Controls smoothness of parameter trajectories or family-switching paths; value chosen to balance fit and stability.

axioms (1)

domain assumption Simplifying assumption: conditional copulas in higher trees depend on conditioning variables only through their marginal conditional distributions.
Invoked to equate the held-out contrast to the higher-tree component of vine total-correlation decomposition at the population level.

pith-pipeline@v0.9.0 · 5596 in / 1505 out tokens · 52294 ms · 2026-05-08T17:10:57.120187+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

55 extracted references · 6 canonical work pages · 1 internal anchor

[1]

and Czado, C

Aas, K. and Czado, C. and Frigessi, A. and Bakken, H. , title =. Insurance: Mathematics and Economics , volume =. 2009 , doi =

2009
[2]

and Cooke, R

Bedford, T. and Cooke, R. M. , title =. Annals of Statistics , volume =. 2002 , doi =

2002
[3]

Brechmann, E. C. and Czado, C. and Aas, K. , title =. Canadian Journal of Statistics , volume =. 2012 , doi =

2012
[4]

Brechmann, E. C. and Joe, H. , title =. Journal of Multivariate Analysis , volume =. 2015 , doi =

2015
[5]

, title =

Akaike, H. , title =. IEEE Transactions on Automatic Control , volume =. 1974 , doi =

1974
[6]

and Solnik, B

Longin, F. and Solnik, B. , title =. Journal of Finance , volume =. 2001 , doi =

2001
[7]

and McNeil, A

Embrechts, P. and McNeil, A. and Straumann, D. , title =. Risk Management: Value at Risk and Beyond , editor =
[8]

Cohen, M. R. and Kohn, A. , title =. Nature Neuroscience , volume =
[9]

Selecting and estimating regular vine copulae and application to financial returns , journal =

Di. Selecting and estimating regular vine copulae and application to financial returns , journal =
[10]

Engle, R. F. , title =. Journal of Business & Economic Statistics , volume =
[11]

and Park, Y

Hallac, D. and Park, Y. and Boyd, S. and Leskovec, J. , title =. Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages =. 2017 , publisher =

2017
[12]

and Hastie, T

Friedman, J. and Hastie, T. and Tibshirani, R. , title =. Biostatistics , volume =. 2008 , doi =

2008
[13]

and Ghanmi, M

Janke, T. and Ghanmi, M. and Steinke, F. , title =. Advances in Neural Information Processing Systems , volume =. 2021 , url =

2021
[14]

and Kraus, D

Killiches, M. and Kraus, D. and Czado, C. , title =. Australian & New Zealand Journal of Statistics , volume =. 2017 , doi =

2017
[15]

and Czado, C

Kreuzer, A. and Czado, C. , title =. arXiv preprint arXiv:1911.00702 , year =. doi:10.48550/arXiv.1911.00702 , url =

work page doi:10.48550/arxiv.1911.00702 1911
[16]

Acar, E. F. and Czado, C. and Lysy, M. , title =. Econometrics and Statistics , volume =
[17]

and Amvrosiadis, T

Kudryashova, N. and Amvrosiadis, T. and Dupuy, N. and Rochefort, N. and Onken, A. , title =. PLoS Computational Biology , volume =
[18]

and Lederer, J

Laszkiewicz, M. and Lederer, J. and Fischer, A. , title =. arXiv preprint arXiv:2107.07352 , year =. doi:10.48550/arXiv.2107.07352 , url =

work page doi:10.48550/arxiv.2107.07352
[19]

Ling, C. K. and Fang, F. and Kolter, J. Z. , title =. Advances in Neural Information Processing Systems , volume =. 2020 , url =

2020
[20]

and Broadrick, O

Liu, A. and Broadrick, O. and Niepert, M. and Van den Broeck, G. , title =. International Conference on Learning Representations , year =
[21]

and Sun, Z

Ma, J. and Sun, Z. , title =. Tsinghua Science and Technology , volume =
[22]

and Doucet, A

Davy, M. and Doucet, A. , title =. IEEE Signal Processing Letters , volume =. 2003 , doi =

2003
[23]

and Steel, M

Huk, D. and Steel, M. and Dutta, R. , title =. Proceedings of The 28th International Conference on Artificial Intelligence and Statistics , pages =. 2025 , editor =

2025
[24]

and Damoulas, T

Huk, D. and Damoulas, T. , title =. International Conference on Learning Representations , year =
[25]

Amortized Vine Copulas for High-Dimensional Density and Information Estimation

Safaai, H. , title =. arXiv preprint arXiv:2604.20568 , year =. doi:10.48550/arXiv.2604.20568 , url =

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.20568
[26]

and Wang, A

Safaai, H. and Wang, A. Y. and Kira, S. and Malerba, S. B. and Panzeri, S. and Harvey, C. D. , title =. Nature Neuroscience , volume =. 2025 , doi =

2025
[27]

and Amvrosiadis, T

Mitskopoulos, L. and Amvrosiadis, T. and Onken, A. , title =. Frontiers in Neuroscience , volume =
[28]

and Onken, A

Mitskopoulos, L. and Onken, A. , title =. Entropy , volume =. 2023 , doi =

2023
[29]

, title =

Nagler, T. , title =. arXiv preprint arXiv:2410.16806 , year =

work page arXiv
[30]

Nagler, T. and Kr. Stationary vine copula models for multivariate time series , journal =
[31]

and Wood, F

Berkes, P. and Wood, F. and Pillow, J. W. , title =. Advances in Neural Information Processing Systems , volume =
[32]

and Panzeri, S

Onken, A. and Panzeri, S. , title =. Advances in Neural Information Processing Systems , volume =. 2016 , url =

2016
[33]

and Onken, A

Safaai, H. and Onken, A. and Harvey, C. D. and Panzeri, S. , title =. Physical Review E , volume =. 2018 , doi =

2018
[34]

Patton, A. J. , title =. International Economic Review , volume =. 2006 , doi =

2006
[35]

Patton, A. J. , title =. Journal of Applied Econometrics , volume =. 2006 , doi =

2006
[36]

and Vatter, T

Nagler, T. and Vatter, T. , title =. 2025 , howpublished =. doi:10.5281/zenodo.14616803 , url =

work page doi:10.5281/zenodo.14616803 2025
[37]

Rezende, D. J. and Mohamed, S. , title =. Proceedings of the 32nd International Conference on Machine Learning , pages =. 2015 , url =

2015
[38]

and Sohl-Dickstein, J

Dinh, L. and Sohl-Dickstein, J. and Bengio, S. , title =. International Conference on Learning Representations , year =
[39]

Dalgleish, H. W. P. and Russell, L. E. and Packer, A. M. and Roth, A. and Gauld, O. M. and Greenstreet, F. and Thompson, E. J. and H. How many neurons are sufficient for perception of cortical activity? , journal =. 2020 , doi =

2020
[40]

Dalgleish, H. W. P. and Russell, L. E. and Packer, A. M. and Roth, A. and Gauld, O. M. and Greenstreet, F. and Thompson, E. J. and H. 20191203\_L541 , howpublished =. 2020 , doi =

2020
[41]

2024 , doi =

Visual Behavior Neuropixels dataset , howpublished =. 2024 , doi =

2024
[42]

and Mediano, P

Rosas, F. and Mediano, P. A. M. and Gastpar, M. and Jensen, H. J. , title =. Physical Review E , volume =
[43]

and Spanhel, F

Schellhase, C. and Spanhel, F. , title =. Statistics and Computing , volume =. 2018 , doi =

2018
[44]

and Berry II, M

Schneidman, E. and Berry II, M. J. and Segev, R. and Bialek, W. , title =. Nature , volume =
[45]

, title =

Sklar, A. , title =. Publications de l'Institut de Statistique de l'Universit
[46]

and Ackerer, D

Tagasovska, N. and Ackerer, D. and Vatter, T. , title =. Advances in Neural Information Processing Systems , volume =. 2019 , url =

2019
[47]

and Ozdemir, F

Tagasovska, N. and Ozdemir, F. and Brando, A. , title =. Proceedings of the 26th International Conference on Artificial Intelligence and Statistics , pages =
[48]

and Vatter, T

Cheng, T. and Vatter, T. and Nagler, T. and Chen, K. , title =. arXiv preprint arXiv:2506.13318 , year =

work page arXiv
[49]

and Schepsmeier, U

Nagler, T. and Schepsmeier, U. and Stoeber, J. and Brechmann, E. C. , title =. 2024 , note =

2024
[50]

Belghazi, M. I. and Baratin, A. and Rajeshwar, S. and Ozair, S. and Bengio, Y. and Courville, A. and Hjelm, R. , title =. Proceedings of the 35th International Conference on Machine Learning , pages =. 2018 , url =

2018
[51]

Claassen, J. N. and Koks, E. E. and de Ruiter, M. C. and Ward, P. J. and J. Journal of Open Source Software , volume =. 2024 , doi =

2024
[52]

and Moroni, M

Panzeri, S. and Moroni, M. and Safaai, H. and Harvey, C. D. , title =. Nature Reviews Neuroscience , volume =
[53]

and Segev, R

Ganmor, E. and Segev, R. and Schneidman, E. , title =. Proceedings of the National Academy of Sciences , volume =
[54]

van der Vaart, A. W. , title =
[55]

Cayco-Gajic, N. A. and Zylberberg, J. and Shea-Brown, E. , title =. Frontiers in Computational Neuroscience , volume =