Research map / Probabilistic and scientific modeling / Dynamical systems
Learning Time-Varying Multi-Region Brain Communications via Scalable Markovian Gaussian Processes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes the Adaptive Delay Model (ADM), a Markovian Gaussian Process / state-space model with continuously time-varying inter-region temporal delays, fit with a parallel-scan Kalman EM for scalable analysis of multi-region spike trains.
Problem
Existing multi-region communication models either ignore temporal delays or assume static or discretely changing delays, and GP-based delay models scale poorly (O(T^3)) or sequentially (O(T)) for long recordings.
Approach
Establishes a universal connection between any temporally stationary GP kernel and a Markovian SSM (via regression-derived transition matrices), letting the MOSE kernel's delay parameter vary over time, and uses a parallel-scan Kalman EM for O(log T) time inference.
Key results
- On V1-V2 spike trains, ADM achieves higher test log-likelihood than MRM-GP and DLAG under the same latent settings, attributed to continuously time-varying delays.
- On five-region Neuropixels data, ADM achieves better test log-likelihood than mDLAG with m_a=3, m_w=1, P=5.
- Parallel scan Kalman EM is significantly faster than sequential EM as T grows (synthetic up to T=600).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Approximate the GP with an SSM | GP regression with Exp kernel · GP regression, Exp | Reg-MSE / 10^-1: 5.7 ± 0.1 -> 5.9 ± 0.1 | Hurt |
| Approximate the GP with an SSM | GP regression with Matern 3/2 kernel · GP regression, Matern 3/2 | Reg-MSE / 10^-1: 5.9 ± 0.2 -> 6.2 ± 0.1 | Hurt |
| Approximate the GP with an SSM | GP regression with SE kernel · GP regression, SE | Reg-MSE / 10^-1: 3.1 ± 0.1 -> 3.3 ± 0.1 | Hurt |
| Approximate the GP with an SSM | GP regression with RQ kernel · GP regression, RQ | Reg-MSE / 10^-1: 3.0 ± 0.1 -> 3.4 ± 0.1 | Hurt |
| Approximate the GP with an SSM | GP regression with SM kernel · GP regression, SM | Reg-MSE / 10^-1: 3.0 ± 0.2 -> 3.3 ± 0.2 | Hurt |
| Approximate the GP with an SSM | GP regression with MOSE kernel · GP regression, MOSE | Reg-MSE / 10^-1: 7.4 ± 0.02 -> 7.6 ± 0.04 | Hurt |
| Approximate the GP with an SSM | GP regression with MOSM kernel · GP regression, MOSM | Reg-MSE / 10^-1: 7.5 ± 0.02 -> 7.9 ± 0.08 | Hurt |
| Approximate the GP with an SSM | GP regression with CSM kernel · GP regression, CSM | Reg-MSE / 10^-1: 8.2 ± 0.05 -> 7.7 ± 0.09 | Helped |
| Approximate the GP with an SSM | GP regression with LMC kernel · GP regression, LMC | Reg-MSE / 10^-1: 0.66 ± 0.02 -> 0.72 ± 0.02 | Hurt |
| Use P=1 instead of P≥2 (ablation) | ADM · Synthetic Data | test log-likelihood | Hurt |
| Vary SSM order P across 2–5 (ablation) | ADM · Synthetic Data | test log-likelihood | No clear effect |
| Vary SSM order P (ablation) | ADM · Two Brain Regions (V1–V2) | validation log-likelihood | Mixed |
Surprising findings
- Their GP-SSM regression comparison shows the SSM approximation has MSE comparable to the exact GP across kernels.
Limitations
- Cubic time complexity in number of regions N and SSM order P.
- Linear-Gaussian observation model; assumes linear/Gaussian latent-to-observation relationship.
Experimental setup
Datasets: Synthetic data V1-V2 monkey spike trains (session 106r001p26) Allen Institute Visual Coding Neuropixels session 750749662
Benchmarks: V1-V2 spike trains Allen Neuropixels five-region spike trains synthetic delay recovery
Metrics: test log-likelihood validation log-likelihood runtime MSE and Pearson correlation of delays
Baselines: DLAG mDLAG MRM-GP mp-srSLDS
Builds on: DLAG mDLAG MRM-GP Markovian Gaussian Processes parallel scan Kalman filter and smoother
Closest papers by meaning
- Multi-Region Markovian Gaussian Process: An Efficient Method to Discover Directional Communications Across Multiple Brain Regions (ICML 2024 · similarity 0.93)
- Accurate Identification of Communication Between Multiple Interacting Neural Populations (ICML 2025 · similarity 0.88)
- One-hot Generalized Linear Model for Switching Brain State Discovery (ICLR 2024 · similarity 0.87)
- Dynamical modeling of nonlinear latent factors in multiscale neural activity with real-time inference (NeurIPS 2025 · similarity 0.86)
- Modeling Latent Neural Dynamics with Gaussian Process Switching Linear Dynamical Systems (NeurIPS 2024 · similarity 0.86)
- CREIMBO: Cross-Regional Ensemble Interactions in Multi-view Brain Observations (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.