Decisive Times - Intrinsic Timescales & Decoding Stability

1. Overview

Recent advances in large-scale neural recordings and computational neuroscience have enabled brain-wide investigations of temporal dynamics underlying cognition. Intrinsic neural timescales (τ) describe how long neural activity remains autocorrelated and have been shown to form a cortical hierarchy, with sensory areas exhibiting short timescales and association cortices exhibiting longer ones (Murray et al., 2014). However, how these intrinsic dynamics relate to decision stability and confidence-like behavior remains unresolved.

This project will use publicly available data from the International Brain Laboratory (IBL), requiring no additional ethics approval. Computational resources are available through institutional high-performance computing facilities.

Timeline: 6 months

M1: Setup & Subset Start: Clean/align IBL sessions; region mapping; movement/pose regression; define epochs. Compute first τ, stability (CTG/trajectories), and DSI on subset brain regions.
Output: QC note + first metrics.
M2: Subset Complete & First Analyses: Finish subset; validate DDM/DSI; run RQ1–RQ3 in parallel (incl. pre/post PCA/GPFA ).
Output: Timescale–Stability Atlas v0, Stability–Behavior maps v0, Retention summary v0.
M3: Scale-out Phase 1: Add well-sampled regions (thresholded by sessions/units); batch τ, stability, DSI; update models.
Output: Atlas v1 (expanded) + CIs.
M4: Robustness & Equivalence: Leave-mouse/lab-out, movement-regressed vs raw, parameter sensitivity, permutation nulls; equivalence tests for any nulls.
Output: Locked analysis config + robustness appendix.
M5: Scale-out Phase 2: Complete regional coverage; finalize mixed-effects for RQ1–RQ3; draft figures.
Output: Atlas v2, regional slopes, retention plots.
M6: Integrate & Package: Final figures (Atlas, Stability–Behavior, Retention), narrative per outcome table, code/README.
Output: Micropublication (≤1,500 words) + 2–3 figures (Timescale–Stability Atlas, Stability–Behavior map, Latent-retention bars) + 1 QC/coverage table + minimal reproducible repo (DOI).

Funding and facilities: This project is supported by Neuromatch Impact Scholars Program. Access to IBL data is confirmed through data sharing agreements. No additional ethics approval is required as all data are publicly available and de-identified. Original data collection was approved by institutional animal care committees at participating IBL labs.

2. Introduction

Decision-making emerges from distributed neural circuits operating across multiple temporal scales. Prior work demonstrated a cortical hierarchy of intrinsic timescales, suggesting that higher-order regions integrate information over longer windows (Murray et al., 2014). These longer timescales are hypothesized to support persistent activity and evidence accumulation.

Intrinsic neural timescales topographically organized along functional connectivity

© 2022 Manea et al., eLife 11:e75540, Fig. 3—figure supplement 1. Reproduced under CC BY 4.0.

Neural stability, defined as the persistence of decision-related representations across time within trials, can be quantified using cross-temporal decoding approaches. Stable neural codes are thought to provide reliable substrates for downstream readout and behavioral commitment. However, the mechanistic link between intrinsic timescales and stability remains largely unexplored at the brain-wide level.

Confidence research in animals has relied on behavioral proxies, such as reaction time, accuracy, and computational modeling of decision variables (Kepecs et al., 2008; Drugowitsch et al., 2012). These approaches suggest that stronger internal decision variables correspond to higher confidence. Yet, it remains unclear whether neural stability itself is a predictor of such decision strength.

Another unresolved question concerns dimensionality. Large-scale recordings reveal high-dimensional neural activity, yet many cognitive variables appear to occupy low-dimensional manifolds (Cunningham & Yu, 2014). If stability and its behavioral relevance are preserved after dimensionality reduction, this would suggest that stability is a robust, low-dimensional feature rather than a fragile high-dimensional artifact.

Research Questions:

RQ1. Are longer intrinsic neural timescales associated with greater temporal stability of decision-related neural activity across different brain regions?

RQ2. Is greater stability of neural activity associated with higher-confidence behaviour, as measured by a decision strength metric?

RQ3. Does compression into lower-dimensional latent dynamics preserve or alter the expression of neural stability and its relation to higher-confidence behavior?

Hypotheses:

H1 – Regional Timescale–Stability Hierarchy

Brain regions will exhibit a hierarchical organization of intrinsic timescales and neural stability during visual decision-making.

H1a (Sensory Regions): Sensory areas (e.g., V1, VISp) will show short intrinsic timescales (log(τ) < 1.8) and low cross-temporal generalization (CTG < 0.65).
H1b (Association/Prefrontal Regions): Association and prefrontal areas (e.g., PPC, MOs) will show long intrinsic timescales (log(τ) > 2.3) and high CTG (CTG > 0.75).
H1c (Intermediate Regions): Intermediate regions (e.g., parietal cortex, striatum) will exhibit intermediate timescales (1.8 < log(τ) < 2.3) and moderate CTG (0.65 < CTG < 0.75).
H1d (Global Relationship): Across regions, intrinsic timescale (log(τ)) will correlate positively with CTG (r > 0.3).

H2 – Stability–Behavior Coupling

Neural stability will predict trial-by-trial decision strength.

H2a (Main Effect): Higher CTG (cross-temporal generalization) values will predict higher Decision Strength Index (DSI), with a positive regression coefficient (β > 0) in mixed-effects models controlling for reaction time and accuracy.
H2b (Moderation by Timescale): The stability–behavior relationship will be stronger in regions with longer intrinsic timescales, reflected by a positive interaction term (log(τ) × CTG → DSI; β_interaction > 0).

H3 – Dimensionality and Stability Preservation

Stability and stability–behavior coupling will be preserved under low-dimensional neural embeddings.

H3a (Stability Retention): PCA-based dimensionality reduction will retain >80% of stability information (CTG_lowD / CTG_fullD > 0.80) within the top 10 principal components.
H3b (Behavioral Coupling Retention): The slope of the stability–behavior relationship will be retained at >70% in low-dimensional space (β_lowD / β_fullD > 0.70).
H3c (Region-Specific Retention): Long-timescale regions (association/prefrontal) will show higher retention (>85%) than sensory regions (<75%), with intermediate regions showing moderate retention (75–85%).

Impact Statement: This project delivers the first brain-wide mapping of the relationship between intrinsic timescales and neural stability across the cortical hierarchy, establishing a mechanistic link between circuit-level temporal dynamics and confidence-like behavior in animal models. By unifying previously separate literatures on intrinsic timescales and decision confidence, it introduces both conceptual and methodological advances, including the Decision Strength Index (DSI)—a DDM/attractor-network-based proxy for confidence in non-human subjects—and the Timescale–Stability Atlas, a novel framework for quantifying regional dynamical organization. More fundamentally, the project addresses whether neural stability constitutes a robust low-dimensional feature of large-scale activity or a high-dimensional artifact, and clarifies how hierarchical temporal structure enables reliable cognition.

Illustration of the drift-diffusion model (DDM)

© 2022 von Bastian, C. C., Reinhartz, A., Udale, R. C., Grégoire, S., Essounni, M., Belleville, S., & Strobach, T., BMC Psychology 10:168. Licensed CC BY 4.0.

3. Methods

Participants / Sample

This study uses publicly available electrophysiological recordings from the International Brain Laboratory (IBL) collected during a standardized visual discrimination task (DOI: 10.6084/m9.figshare.11636748). No new data collection will be performed.

Inclusion criteria:

≥200 trials per session
≥50 well-isolated units per region
Behavioral accuracy >65%
ISI violations <0.5%, amplitude cutoff <0.1

Exclusion criteria:

Probe drift >100 μm
Regions with <3 sessions across mice
Reaction times <100 ms or >5 s

Target sample:

250 sessions from ~60 mice across 8 labs
15 brain regions
25,000 neurons and >60,000 trials

Materials and Procedures

Behavioral task:

Mice perform a two-alternative forced choice (2AFC) visual contrast discrimination task. A Gabor stimulus appears on the left or right at varying contrast levels (0–100%). Mice report stimulus location by rotating a wheel; correct responses are rewarded. Trials include stimulus presentation, response, and 1–2 s inter-trial interval.

Neural recordings:

Simultaneous multi-region recordings were obtained using Neuropixels 1.0 probes (384 channels). Spike sorting was performed with the standardized IBL pipeline (Kilosort2 + manual curation). Regions span sensory (VISp), associative (PPC, frontal cortex), striatal, thalamic, midbrain, and hippocampal areas.

All analyses are conducted on the public IBL dataset with open and reproducible workflows.

Planned Analyses

Module 1: Intrinsic Timescale Estimation

Objective: Quantify intrinsic neural timescales (τ) across brain regions to test the hierarchical organization predicted in H1.

Procedure:

Bin spike trains at Δt = 25 ms
Compute autocorrelation using improved Spike Time Tiling Coefficient (iSTTC)
Fit single-exponential decay model: y(t) = a·e^(-bt) + c
Extract intrinsic timescale: τ = 1/b
Compute per-session, per-region median log(τ)

Hypothesis-Specific Predictions:

H1a: Sensory regions (V1, VISp) will yield log(τ) < 1.8
H1b: Association/prefrontal regions (PPC, MOs) will yield log(τ) > 2.3
H1c: Intermediate regions (parietal, striatum) will yield 1.8 < log(τ) < 2.3
H1d: Across regions, log(τ) will correlate positively with CTG (r > 0.3)

Statistical Test: Mixed-effects linear regression: CTG ~ log(τ) + (1|mouse) + (1|session)

Module 2: Neural Stability Quantification

Objective: Measure neural stability via cross-temporal generalization (CTG) to test regional hierarchy (H1a-c) and its relationship with timescales (H1d).

Cross-Temporal Generalization (CTG):

Train linear SVM classifier on choice (left vs. right) at each time point t₁ during decision epoch (-200 to +500 ms relative to stimulus onset)
Test classifier on all other time points t₂
Compute time-time AUC matrix
Stability = mean off-diagonal AUC within ±300 ms window
Use 5-fold cross-validation

Hypothesis-Specific Predictions:

H1a: Sensory regions will show CTG < 0.65
H1b: Association/prefrontal regions will show CTG > 0.75
H1c: Intermediate regions will show 0.65 < CTG < 0.75
H1d: log(τ) and CTG will correlate positively (r > 0.3)

Statistical Tests:

One-sample t-tests comparing region-averaged CTG against thresholds (0.65, 0.75)
Pearson correlation between log(τ) and CTG across region-session pairs
FDR correction for multiple regions

Robustness checks:

Time–time representational similarity analysis (RSA) using correlation distance
Trajectory dwell-time metrics (time to diverge from initial state)

Module 3: Decision Strength Modeling

Objective: Derive trial-by-trial Decision Strength Index (DSI) to test stability–behavior coupling (H2a) and timescale-modulated coupling (H2b).

Hierarchical Drift-Diffusion Model (DDM):

Drift rate (v) varies with stimulus strength
Boundary separation (a), non-decision time (t₀), and lapse rate (π) estimated per mouse
Trial-level latent decision variable magnitude = |evidence at boundary|
Decision Strength Index (DSI) = normalized decision variable magnitude

Model fitting: Bayesian hierarchical modeling using Stan (MCMC sampling)

Validation:

RT-conditioned accuracy curves
Lapse rate controls (<10% per mouse)
Posterior predictive checks
Leave-one-mouse-out cross-validation

Hypothesis-Specific Predictions:

H2a: β_CTG > 0 in mixed-effects model: DSI ~ CTG + RT + Accuracy + (1|mouse) + (1|session)
H2b: β_interaction > 0 in extended model: DSI ~ CTG × log(τ) + CTG + log(τ) + RT + Accuracy + (1|mouse) + (1|session)

Statistical Tests:

Mixed-effects regression for H2a (primary test)
Interaction term significance for H2b
Region-stratified analyses to identify which regions drive stability–behavior coupling

Module 4: Dimensionality Reduction

Objective: Test whether stability (H3a), stability–behavior relationships (H3b), and region-specific compression profiles (H3c) are preserved in low-dimensional neural dynamics.

Methods:

Principal Component Analysis (PCA)
Gaussian Process Factor Analysis (GPFA)

Procedure:

Train dimensionality reduction on 70% of trials
Test on held-out 30%
Compute stability (CTG) in low-D space (top 5, 10, 20 components)
Re-fit stability–behavior model (DSI ~ CTG_lowD + covariates)
Calculate retention metrics:
- Stability retention = (CTG_lowD / CTG_fullD) × 100%
- Slope retention = (β_lowD / β_fullD) × 100%

Hypothesis-Specific Predictions:

H3a: Top 10 PCs will retain >80% of stability (CTG_lowD / CTG_fullD > 0.80)
H3b: Top 10 PCs will retain >70% of stability–behavior slope (β_lowD / β_fullD > 0.70)
H3c:
- Long-τ regions (association/prefrontal): retention >85%
- Short-τ regions (sensory): retention <75%
- Intermediate-τ regions: retention 75-85%

Statistical Tests:

One-sample t-tests against retention thresholds (H3a, H3b)
ANOVA comparing retention rates across region groups (H3c)
Post-hoc comparisons: sensory vs. association vs. intermediate

Statistical Analysis Summary

Primary hypothesis tests:

H1a-c: One-sample t-tests and region-group comparisons for timescale and CTG thresholds
H1d: Mixed-effects regression: CTG ~ log(τ) + (1|mouse) + (1|session)
H2a: Mixed-effects regression: DSI ~ CTG + RT + Accuracy + (1|mouse) + (1|session)
H2b: Mixed-effects regression with interaction: DSI ~ CTG × log(τ) + CTG + log(τ) + RT + Accuracy + (1|mouse) + (1|session)
H3a-b: One-sample t-tests against retention thresholds (80%, 70%)
H3c: ANOVA on retention rates across region groups + post-hoc pairwise comparisons

Multiple comparisons correction: False Discovery Rate (FDR) correction applied across 15 brain regions for regional analyses (H1a-c, H3c)

Ethics and Data Management

Ethics: This study uses publicly available, de-identified data from the International Brain Laboratory. No new animal procedures are performed. Original data collection was approved by institutional animal care and use committees at all participating IBL laboratories (see IBL data release documentation).

Data sharing: All analysis code will be made available on GitHub under MIT license. Processed data matrices (timescales, stability metrics, DDM parameters) will be archived on Open Science Framework (OSF). Interactive visualizations will be hosted via Jupyter notebooks. Data will be released upon publication or preprint submission, whichever comes first.

Confidentiality: Not applicable (public dataset, no human subjects).

BUDGET:

Category	Description	Amount (USD)
Cloud Computing	High-performance cloud resources for large-scale decoding, modeling, and storage	$1,500
Conference Travel	Registration, travel, and accommodation for presentation at one international neuroscience conference	$2,500
Publication Costs	Open-access publication fees	$1,500
Contingency	Buffer for unforeseen computational or dissemination costs	$1,000
Total		$6,500

References
Murray, J.D., Bernacchia, A., Freedman, D.J., et al. (2014). A hierarchy of intrinsic timescales across primate cortex. Nature Neuroscience, 17:1661–1663. https://doi.org/10.1038/nn.3862.

Kepecs, A., Uchida, N., Zariwala, H.A., Mainen, Z.F. (2008). Neural correlates, computation and behavioural impact of decision confidence. Nature, 455:227–231. https://doi.org/10.1038/nature07200.

Drugowitsch, J., Moreno-Bote, R., Churchland, A.K., Shadlen, M.N., Pouget, A. (2012). The cost of accumulating evidence in perceptual decision making. Journal of Neuroscience, 32:3612–3628. https://doi.org/10.1523/JNEUROSCI.4010-11.2012.

Cunningham, J.P., Yu, B.M. (2014). Dimensionality reduction for large-scale neural recordings. Nature Neuroscience, 17:1500–1509. https://doi.org/10.1038/nn.3776.

Damaraju, E., Allen, E.A., Belger, A., et al. (2014). Dynamic functional connectivity analysis reveals transient states of dysconnectivity in schizophrenia. NeuroImage: Clinical, 5:298–308. https://doi.org/10.1016/j.nicl.2014.07.003.

Luo, L., Li, Q., You, W., et al. (2021). Altered brain functional network dynamics in obsessive-compulsive disorder. Human Brain Mapping, 42:2061–2076. https://doi.org/10.1002/hbm.25345.

Authors & ResearchHub Profiles

Zhiyi Yang: https://www.researchhub.com/author/8561649

Vikramaditya Bisani: https://www.researchhub.com/author/8561648

Süheyl Berat Gülenç: https://www.researchhub.com/author/8561650

Kayla Stafford: https://www.researchhub.com/author/8556105

Jason Manley (supervisor): https://www.researchhub.com/author/8562208

Decisive Times - Intrinsic Timescales & Decoding Stability

Decisive Times - Intrinsic Timescales & Decoding Stability

Funding Progress

Related Funding

Request for Proposals - Neuroscience

Nonprofit Recipient

Topics

DOI