IntrinsicTimescales.jl: A Julia package to estimate intrinsic (neural) timescales (INTs) from time-series data

Georg Northoff; Yasir Catal

REVIEW 4 major objections 6 minor 22 references

IntrinsicTimescales.jl: A Julia package to estimate intrinsic (neural) timescales (INTs) from time-series data

T0 review · 4 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read IntrinsicTimescales.jl bundles eight ways to estimate neural timescales in one Julia package.

desk verdict A useful software-tool preprint whose central claim—that the package actually works—rests entirely on the authors' word, since no code, repository, or validation is included. read the letter →

arxiv 2505.11507 v1 pith:2FUBJG3G submitted 2025-05-02 q-bio.NC physics.data-an

Yasir Catal , Georg Northoff This is my paper · ORCID

classification q-bio.NCphysics.data-an

keywords intrinsicneuraltimescalesJuliapackageautocorrelationfunctionpowerspectraldensityOrnstein-UhlenbeckprocessapproximateBayesiancomputationvariationalinferencemissingdata

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

IntrinsicTimescales.jl is a Julia package that collects the standard and newer ways of estimating intrinsic neural timescales (INTs), the time window over which a brain region's past activity shapes its response to new input. The central claim is that the package gives researchers one standardized, high-performance tool that computes six model-free INT estimates from the autocorrelation function or power spectrum, plus two Bayesian estimates from an Ornstein-Uhlenbeck process fit to the data. This matters because INTs are increasingly used in cognitive and clinical neuroscience, while existing software is scattered across methods, often lacks uncertainty quantification, and does not cleanly handle missing data or oscillatory artifacts. If the package works as described, a researcher can estimate and compare all these measures on the same recordings, with posterior-predictive checks for the Bayesian routes.

What carries the argument

The central object is the package's unified estimation interface, backed by two families of mechanisms. For the model-free methods, the machinery is the autocorrelation function (ACF) and the power spectral density (PSD): the ACF is computed by inverse Fourier transform of the squared Fourier magnitude when the data are complete and by a direct time-domain sum that respects missingness when they are not; the PSD is computed as a windowed periodogram, or with a spectral estimator for unevenly sampled data when samples are missing. The knee estimator relies on the Fourier relation between ACF decay and spectral knee, plus a Lorentzian-with-peaks fitting procedure from reference [6]. For the Bayesian methods, the load-bearing model is the Ornstein-Uhlenbeck process, a mean-reverting random process whose decay constant is taken to be the intrinsic timescale; adaptive approximate Bayesian computation and automatic-differentiation variational inference give two ways to form a posterior over that decay constant, and the generative model itself is what lets the package handle missing data and oscillatory artifacts without discarding recordings.

What would settle it

Take a long Ornstein-Uhlenbeck simulation with known decay time $\tau$, mask about 30 percent of the samples to imitate data loss, and run both Bayesian estimators; if either posterior fails to concentrate near the true $\tau$, or the two methods disagree sharply, the package's claim to deliver unbiased timescale estimates would be contradicted.

Watch

Extended reading notes

Core claim

The paper's central claim is that IntrinsicTimescales.jl is a complete, modular toolbox for estimating intrinsic neural timescales from time-series data. On the model-free side, it computes the autocorrelation function by inverse Fourier transform of the squared spectrum when data are complete, or directly in the time domain when they are not, and derives from it ACW-50, ACW-0, ACW-e, AUC, and tau; the knee estimator instead fits a Lorentzian to the power spectral density and converts the knee frequency into a decay rate through the Fourier relation between autocorrelation and spectrum, following a three-step peak-cleaning procedure from reference [6]. On the Bayesian side, the package estimates the decay constant of an Ornstein-Uhlenbeck process using adaptive approximate Bayesian computation and automatic-differentiation variational inference, with the generative model extended to cover missing samples and oscillatory artifacts. Plotting utilities and posterior-predictive checks complete the package. The claim is that this fills the gap of a missing standardized tool for INT estimation.

Load-bearing premise

The Bayesian estimates stand on the assumption that each neural time series behaves like a single Ornstein-Uhlenbeck process, a random walk pulled back toward its baseline, whose decay constant is the intrinsic timescale, and the paper does not test whether real brain signals satisfy that model.

Editorial extensions

If this is right

A single script can now produce ACW-50, ACW-0, ACW-e, AUC, tau, knee, and two Bayesian posterior estimates from the same recording, making cross-method comparison routine instead of a software-integration project.
Recordings with missing segments or oscillatory contamination, common in clinical and scalp-EEG data, can still yield timescale estimates with quantified uncertainty rather than being discarded.
Posterior-predictive plots give users a direct visual check of whether the Ornstein-Uhlenbeck model is appropriate for a particular signal before interpreting the estimated timescale.
The paper argues that the Julia implementation keeps Bayesian inference tractable on large neuroimaging datasets where slower interpreted-language tools would be impractical.
A Python frontend is planned, which should make the same estimators accessible to researchers whose pipelines are built outside Julia.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Inference: because the package exposes several estimators through one interface, any systematic disagreement among them on real data will become visible; turning that disagreement into a benchmark is a natural next step the paper does not take.
Inference: the two Bayesian estimators should agree on data that genuinely follow an Ornstein-Uhlenbeck process; where they diverge, the divergence points to model misspecification or numerical issues rather than to a brain property.
Inference: the same estimators transfer to non-neural time series, such as climate or financial records, since nothing in the mathematical core is specific to the brain; that application is outside the paper's scope.
Inference: a systematic simulation sweep over recording length, missing fraction, and oscillation strength would map each estimator's bias and could give principled guidance for choosing among the eight methods; the package contains the machinery to run such a sweep.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

If you need a quick read: this is a software-description preprint with no software attached. The idea is sensible and the feature list is genuinely useful for the INT community—unified access to ACW variants, AUC, tau, knee, and two Bayesian OU-based estimators (aABC and ADVI) with missing-data handling is a real gap. The novelty is the integration, not the estimators, and the authors are clear about that. The statement of need is well-grounded in the literature, and the citation pattern looks honest.

The problem is that the central claim—"IntrinsicTimescales.jl is a high-performance, rigorously tested toolbox"—is unsupported in the manuscript. There is no repository link, no commit hash, no unit tests, no simulation-based validation, no comparison against existing implementations, and no ground-truth recovery experiment. The only URL points to a future Python frontend. The sentence about "rigorous testing" is self-report. For a software paper, that's the load-bearing wall. Several estimators involve nontrivial numerics (inverse-FFT-based ACF, Lomb-Scargle periodogram under missingness, exponential/Lorentzian fitting), and any implementation bug would silently invalidate downstream estimates. The stress-test note is right: even if the Bayesian OU model is exactly what the field wants, we have no evidence the package estimates its parameters correctly.

On the Bayesian side, the single-OU assumption is introduced without discussion. That's a modeling choice, not a fatal flaw—but it deserves at least a paragraph on when it's appropriate and what it means if the data have multiple timescales or nonstationarity.

What's good: the feature list is precise, the methods are described consistently with their cited sources, and the package fills a real practical need. The paper does not overclaim theory; it claims software, and that's exactly what needs proof.

Verdict: this should go to peer review, but only with the strong expectation that the authors provide the repository, a commit hash, and simulation-based validation recovering known timescales, plus some comparison against existing tools. If the code is real and tested, this becomes a practical contribution many labs will use. If not, it's an abstract for a package that may never ship. Recommend conditional acceptance contingent on the artifact and validation. I would not cite it for the package itself until I can run it, but I would cite it as a reference point for the feature landscape.

Referee Report

4 major / 6 minor

Summary. This manuscript announces IntrinsicTimescales.jl, a Julia package designed to estimate intrinsic neural timescales (INTs) from time-series data. The package would implement model-free estimators (ACW-50, ACW-0, ACW-e, AUC, tau, knee) and Bayesian estimators based on an Ornstein-Uhlenbeck process using aABC and ADVI, with specialized handling for missing data and oscillatory artifacts. The paper describes the stated need for such a unified toolbox, lists the package's features, and outlines the documentation structure. No code, repository link, validation results, or usage examples are provided in the manuscript.

Significance. If the package is real, functional, and correct, it would fill a genuine gap: existing INT estimation tools are scattered across languages and often lack uncertainty quantification or missing-data support. A unified, Julia-based toolbox with both model-free and Bayesian estimators, plus posterior predictive checks, would be a useful contribution to the neuroscience community. However, the manuscript as submitted provides no artifact or evidence that the package works, so the significance is prospective rather than demonstrated. The paper does not include machine-checked proofs, reproducible code, parameter-free derivations, or falsifiable predictions; the strengths it claims (high performance, rigorous testing, comprehensive documentation) are assertions only.

major comments (4)

[Major Features] The manuscript provides no access to the package artifact: there is no repository URL, DOI, commit hash, or version for IntrinsicTimescales.jl itself. The only URL cited in the text points to INTpy, a future Python frontend. The central claim that the package exists and estimates INTs cannot be verified by any reader. Please provide a permanent, versioned link to the package repository and the specific version described.
[Statement of Need] The assertion of 'rigorous testing and documentation' is self-reported and unsupported by any evidence in the manuscript. There are no unit tests, no synthetic-data experiments, no ground-truth recovery checks, and no comparisons against existing reference implementations such as Zeraati et al. (2022) or FOOOF. Since the correctness of the estimators is the central claim of a software paper, add a validation section that simulates OU processes and other ground-truth signals with known timescales and reports bias, root-mean-square error, and coverage for every estimator.
[Major Features, knee estimator] The description of the knee procedure is ambiguous. It states that a Lorentzian is fitted to the PSD, then subtracted; gaussians for oscillatory peaks are fitted to the residual; then the gaussians are subtracted from the base PSD and a Lorentzian is fitted 'one more time.' This does not match the FOOOF procedure cited (Gao et al., 2020), where the aperiodic component (including the knee) and the Gaussian peaks are estimated jointly within a single model. Specify the exact objective function, the parameters being fitted in each step, and how the final timescale is derived. Provide a worked example with data.
[Major Features, Bayesian Parameter Estimation] The Bayesian estimators treat the intrinsic timescale as the decay constant of a single Ornstein-Uhlenbeck process, but the manuscript provides no justification that real cortical signals satisfy this generative model. If the model is misspecified, the posterior timescale is a property of the model, not of the brain. At minimum, state this assumption explicitly as a limitation and include posterior predictive checks on surrogate or real data that demonstrate whether the model captures the observed ACF/PSD.

minor comments (6)

[Major Features, ACW-e bullet] The phrase 'time to each 1/e' should be 'time to reach 1/e'.
[Figure 1] Figure 1 is referenced in the text but not included or displayed in the manuscript. Ensure the actual figure appears in the final version.
[Major Features] The sentence 'A diagram shows the major features of the package can be seen in Figure 1' is grammatically awkward; rephrase to something like 'Figure 1 shows the major features of the package.'
[Abstract / Major Features] The term 'novel techniques' for aABC and ADVI is misleading because both methods are already published (Beaumont et al., 2009; Kucukelbir et al., 2017) and have been applied to OU-based timescale estimation in Zeraati et al. (2022). The novelty is in the implementation, not the methodology.
[Documentation] The Documentation section states that all documentation is hosted online but does not provide a URL. State the documentation URL explicitly.
[Abstract] The abstract gives a conceptual definition of INTs as 'the time window in which prior information from an ongoing stimulus can affect the processing of newly arriving information,' but the package's operational definitions (e.g., ACF decay, knee frequency) should be clearly distinguished from this conceptual definition in the main text.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is a software/package description that defines estimators via external references and contains no derivation whose conclusion is equivalent to its inputs.

full rationale

IntrinsicTimescales.jl is a software-description paper rather than a formal derivation chain. Its target quantities are introduced through standard external references: ACW-50/0/e, AUC, tau, and knee are defined as properties of the ACF or PSD, and the Bayesian estimators are presented as implementations of an Ornstein-Uhlenbeck generative model with aABC (Beaumont et al. 2009; Zeraati et al. 2022) and ADVI (Kucukelbir et al. 2017). No quantity is fitted to a subset of data and then 'predicted' in a closely related form; no equation defines an estimator in terms of its own output; and no load-bearing step invokes a self-citation as proof of correctness or uniqueness. The few self-citations (refs. 14 and 15, involving the current authors) are prior empirical findings about intrinsic neural timescales, not evidence used to justify the package's estimators. The 'knee' method explicitly follows the FOOOF three-step procedure with external attribution, so it is an implementation choice rather than a circular re-derivation. Concerns that 'rigorous testing and documentation' is unverified or that no repository link is given are reproducibility/correctness risks, not circularity. Consequently, the appropriate finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new fitted constants or theoretical entities; all estimators come from prior literature. The ledger therefore captures only the background statistical and modeling assumptions the package relies on.

assumptions (3)

standard math Wiener-Khinchine theorem implies that the inverse decay rate of an exponential ACF is proportional to the knee frequency of the PSD.
Invoked in the knee estimator description in Major Features; accepted mathematical result, not derived in the paper.
domain assumption Neural time series can be represented as an Ornstein-Uhlenbeck process after handling oscillations and missing data.
Assumed by both Bayesian estimators in Major Features; not tested in the preprint.
domain assumption ACF computed as the inverse Fourier transform of the squared magnitude of the Fourier transform gives a valid autocorrelation estimate for complete data, and Lomb-Scargle periodograms provide valid spectra with missing data.
The specialized ACF/PSD section relies on these equivalences; boundary, windowing, and missingness effects are not analyzed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of IntrinsicTimescales.jl: A Julia package to estimate intrinsic (neural) timescales (INTs) from time-series data." pith.science (2026). https://pith.science/paper/2FUBJG3G

@misc{pith2026250511507,
  author       = {Pith},
  title        = {Pith review of: IntrinsicTimescales.jl: A Julia package to estimate intrinsic (neural) timescales (INTs) from time-series data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2FUBJG3G}},
  note         = {Machine review of arXiv:2505.11507}
}

read the original abstract

IntrinsicTimescales.jl is a Julia package to perform estimation of intrinsic neural timescales (INTs). INTs are defined as the time window in which prior information from an ongoing stimulus can affect the processing of newly arriving information. INTs are estimated either from the autocorrelation function (ACF) or the power spectral density (PSD) of time-series data. In addition to the model-free estimates of INTs, IntrinsicTimescales.jl offers implementations of novel techniques of timescale estimation via performing parameter estimation of an Ornstein-Uhlenbeck process with adaptive approximate Bayesian computation (aABC) and automatic differentiation variational inference (ADVI).

Figures

Figures reproduced from arXiv: 2505.11507 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

[1]

Hasson, U., Yang, E., Vallines, I., Heeger, D. J. & Rubin, N. A Hierarchy of Temporal Receptive Windows in Human Cortex. J. Neurosci. 28, 2539–2550 (2008)

work page 2008
[2]

& Honey, C

Hasson, U., Chen, J. & Honey, C. J. Hierarchical process memory: memory as an integral component of information processing. Trends Cogn. Sci. 19, 304–313 (2015)

work page 2015
[3]

Wolff, A. et al. Intrinsic neural timescales: temporal integration and segregation. Trends Cogn. Sci. 26, 159–173 (2022)

work page 2022
[4]

Golesorkhi, M. et al. The brain and its time: intrinsic neural timescales are key for input processing. Commun. Biol. 4, 1–16 (2021)

work page 2021
[5]

Honey, C. J. et al. Slow Cortical Dynamics and the Accumulation of Information over Long Timescales. Neuron 76, 423–434 (2012)

work page 2012
[6]

L., Pfeffer, T

Gao, R., van den Brink, R. L., Pfeffer, T. & Voytek, B. Neuronal timescales are functionally dynamic and shaped by cortical microarchitecture. eLife 9, e61277 (2020)

work page 2020
[7]

A., Cornuet, J.-M., Marin, J.-M

Beaumont, M. A., Cornuet, J.-M., Marin, J.-M. & Robert, C. P. Adaptive approximate Bayesian computation. Biometrika 96, 983–990 (2009)

work page 2009
[8]

Zeraati, R., Engel, T. A. & Levina, A. A flexible Bayesian framework for unbiased estimation of timescales. Nat. Comput. Sci. 2, 193–204 (2022)

work page 2022

Show all 22 references

[9]

& Blei, D

Kucukelbir, A., Tran, D., Ranganath, R., Gelman, A. & Blei, D. M. Automatic Differentiation Variational Inference. J. Mach. Learn. Res. 18, 1–45 (2017)

2017
[10]

Murray, J. D. et al. A hierarchy of intrinsic timescales across primate cortex. Nat. Neurosci. 17, 1661– 1663 (2014)

2014
[11]

Macroscopic gradients of synaptic excitation and inhibition in the neocortex

Wang, X.-J. Macroscopic gradients of synaptic excitation and inhibition in the neocortex. Nat. Rev. Neurosci. 21, 169–178 (2020)

2020
[12]

Ito, T., Hearne, L. J. & Cole, M. W. A cortical hierarchy of localized and distributed processes revealed via dissociation of task activations, connectivity changes, and intrinsic timescales. NeuroImage 221, 117141 (2020)

2020
[13]

Zeraati, R. et al. Intrinsic timescales in the visual cortex change with selective attention and reflect spatial connectivity. Nat. Commun. 14, 1858 (2023)

2023
[14]

Çatal, Y. et al. Flexibility of intrinsic neural timescales during distinct behavioral states. Commun. Biol. 7, 1667 (2024)

2024
[15]

Wolman, A. et al. Intrinsic neural timescales mediate the cognitive bias of self - temporal integration as key mechanism. NeuroImage 268, 119896 (2023)

2023
[16]

Zilio, F. et al. Are intrinsic neural timescales related to sensory processing? Evidence from abnormal behavioral states. NeuroImage 226, 117579 (2021)

2021
[17]

& Northoff, G

Golesorkhi, M., Gomez-Pilar, J., Tumati, S., Fraser, M. & Northoff, G. Temporal hierarchy of intrinsic neural timescales converges with spatial core-periphery organization. Commun. Biol. 4, 1–14 (2021)

2021
[18]

Cusinato, R. et al. Intrinsic Neural Timescales in the Temporal Lobe Support an Auditory Processing Hierarchy. J. Neurosci. 43, 3696–3707 (2023)

2023
[19]

Manea, A. M. G. et al. Neural timescales reflect behavioral demands in freely moving rhesus macaques. Nat. Commun. 15, 2151 (2024)

2024
[20]

& Gollo, L

Wu, K. & Gollo, L. L. Mapping and modeling age-related changes in intrinsic neural timescales. Commun. Biol. 8, 1–16 (2025)

2025
[21]

& Masuda, N

Watanabe, T., Rees, G. & Masuda, N. Atypical intrinsic neural timescale in autism. eLife 8, e42256 (2019)

2019
[22]

Raut, R. V. et al. Organization of Propagated Intrinsic Brain Activity in Individual Humans. Cereb. Cortex 30, 1716–1734 (2020)

2020

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

Hasson, U., Yang, E., Vallines, I., Heeger, D. J. & Rubin, N. A Hierarchy of Temporal Receptive Windows in Human Cortex. J. Neurosci. 28, 2539–2550 (2008)

work page 2008

[2] [2]

& Honey, C

Hasson, U., Chen, J. & Honey, C. J. Hierarchical process memory: memory as an integral component of information processing. Trends Cogn. Sci. 19, 304–313 (2015)

work page 2015

[3] [3]

Wolff, A. et al. Intrinsic neural timescales: temporal integration and segregation. Trends Cogn. Sci. 26, 159–173 (2022)

work page 2022

[4] [4]

Golesorkhi, M. et al. The brain and its time: intrinsic neural timescales are key for input processing. Commun. Biol. 4, 1–16 (2021)

work page 2021

[5] [5]

Honey, C. J. et al. Slow Cortical Dynamics and the Accumulation of Information over Long Timescales. Neuron 76, 423–434 (2012)

work page 2012

[6] [6]

L., Pfeffer, T

Gao, R., van den Brink, R. L., Pfeffer, T. & Voytek, B. Neuronal timescales are functionally dynamic and shaped by cortical microarchitecture. eLife 9, e61277 (2020)

work page 2020

[7] [7]

A., Cornuet, J.-M., Marin, J.-M

Beaumont, M. A., Cornuet, J.-M., Marin, J.-M. & Robert, C. P. Adaptive approximate Bayesian computation. Biometrika 96, 983–990 (2009)

work page 2009

[8] [8]

Zeraati, R., Engel, T. A. & Levina, A. A flexible Bayesian framework for unbiased estimation of timescales. Nat. Comput. Sci. 2, 193–204 (2022)

work page 2022