| name | bayesian-ippm-cortical-entrainment |
| description | Bayesian framework for Information Processing Pathway Maps (IPPMs) to map cortical entrainment from EEG/MEG data. Compares Bayesian vs frequentist approaches for model adjudication in computational neuroscience. Uses temporal response functions (TRFs) and model evidence for reconstructing sensory processing pathways. Activation: IPPM, cortical entrainment, Bayesian model comparison, information processing pathway, temporal response function, auditory processing, EEG MEG analysis, model adjudication. |
| metadata | {"arxiv_id":"2607.06284","published":"2026-07-07","authors":"Kaibo Zhang, Ji Wu, Chao Zhang, Andrew Thwaites","tags":["computational-neuroscience","Bayesian-inference","cortical-entrainment","IPPM","EEG","MEG","auditory-processing","model-comparison"]} |
Bayesian IPPM for Cortical Entrainment Mapping
Overview
Novel Bayesian framework for constructing Information Processing Pathway Maps (IPPMs) — formalized representations of the sequence of computational transformations sensory information undergoes in the brain. Provides direct probabilistic comparison between competing computational models of neural processing, replacing traditional frequentist null hypothesis testing.
Core Concepts
Information Processing Pathway Maps (IPPMs)
- Formal representations of sequences of mathematical transformations applied to sensory stimuli
- Map the latency and cortical expression of computational steps
- Generated from high-temporal-resolution neuroimaging (EEG/MEG)
- Serve as scalable tools for reverse-engineering brain processes
Key Innovation: Bayesian Model Adjudication
Traditional (Frequentist) Approach:
- Tests null hypothesis: model output is uncorrelated with neural activity
- Reports P(data | model is irrelevant) — conceptually indirect
- Binary accept/reject decisions at each processing stage
Bayesian Approach:
- Computes P(model is correct | data) — direct probabilistic statement
- Quantifies relative evidence for competing computational hypotheses
- Uses Bayes factors for model comparison
- Naturally handles collinear models through model evidence integration
Methodology
1. Temporal Response Functions (TRFs)
- Linear mapping from stimulus features to neural responses
- Capture latency-specific entrainment patterns
- Each computational model generates predicted TRFs
- Compare predicted vs observed TRFs in EEG/MEG data
2. Bayesian Framework
P(Model_i | Data) ∝ P(Data | Model_i) × P(Model_i)
- Likelihood: P(Data | Model_i) — how well model predicts neural data
- Prior: P(Model_i) — prior belief about model plausibility
- Posterior: P(Model_i | Data) — updated belief after observing data
- Bayes Factor: BF_12 = P(Data|Model_1) / P(Data|Model_2) — relative evidence
3. Model Comparison Workflow
- Define candidate computational models of sensory processing
- Generate TRF predictions for each model
- Fit models to EEG/MEG data using both frequentist and Bayesian approaches
- Compare model evidence across competing hypotheses
- Construct IPPM nodes based on strongest evidence (not just significance)
4. Experimental Validation
- Dataset: Auditory neuroimaging (MEG/EEG)
- Task: Reconstruct known loudness-processing pathway
- Comparison: Bayesian vs frequentist recovery of pathway stages
- Metrics: Pathway reconstruction accuracy, handling of collinear models, evidence accumulation robustness
Key Advantages of Bayesian IPPM
- Collinear Model Handling: Naturally handles correlated predictors through model evidence integration rather than arbitrary threshold selection
- Robust Evidence Accumulation: Bayesian updating allows progressive refinement as more data becomes available
- Quantitative Model Comparison: Continuous measure of relative evidence rather than binary significant/non-significant
- Prior Integration: Incorporates domain knowledge about plausible processing architectures
- Uncertainty Quantification: Full posterior distributions over pathway parameters
Pitfalls
- Prior sensitivity: Results can be sensitive to choice of priors, especially with limited data
- Model misspecification: If true model is not in candidate set, posterior may be overconfident in "least-bad" option
- Computational cost: Bayesian inference (MCMC, variational methods) more expensive than frequentist tests
- Interpretation: Bayes factors require careful calibration — BF > 3 is "substantial" but context-dependent
Applications
- Auditory processing pathway mapping: Reconstruct stages from cochlea to auditory cortex
- Visual processing hierarchies: Map V1 → V2 → V4 → IT transformations
- Language processing: Identify N400/P600 computational sources
- Clinical applications: Compare pathway disruptions in patient populations
- Model validation: adjudicate between competing theories of sensory coding
Implementation Notes
- TRF estimation: Use mTRF toolbox (Crosse et al.) or custom ridge regression
- Bayesian model comparison: Use bridge sampling, thermodynamic integration, or WAIC
- Software: Stan, PyMC, or custom variational inference
- Cross-validation: Essential for assessing generalization of pathway maps
Related Work
- Thwaites et al. (2025) — Original IPPM framework with frequentist statistics
- Crosse et al. (2016) — mTRF toolbox for temporal response functions
- Mesite et al. — Auditory cortex computational models
- Bayesian model selection in neuroscience (Penny et al., 2004)
Activation Keywords
IPPM, cortical entrainment, Bayesian model comparison, information processing pathway, temporal response function, TRF, auditory processing, EEG analysis, MEG analysis, model adjudication, Bayes factor, sensory processing pathway, computational neuroscience, neural encoding model