Bayesian Information Processing Pathway Maps

---
name: bayesian-information-processing-pathway-maps
description: >
  Bayesian framework for quantifying neural entrainment evidence in Information Processing 
  Pathway Maps (IPPMs). Shifts from frequentist hypothesis testing to probabilistic model 
  adjudication, enabling relative evidence quantification for competing computational hypotheses.
  Applied to auditory neuroimaging for reconstructing cortical processing pathways.
tags: [computational-neuroscience, bayesian-inference, neural-entrainment, model-selection,
       auditory-processing, information-pathway, frequentist-vs-bayesian, neuroimaging]
related_skills: [neural-dynamics-analysis-methodology, computational-neuroscience-in-llm-era,
                 bayesian-model-selection-bb-plot]
source: arXiv:2607.06284v1
date: 2026-07-08
---

# Bayesian Information Processing Pathway Maps

## Paper Summary

**Title**: Quantifying Entrainment Evidence: A Comparison of Frequentist and Bayesian Approaches for Information Processing Pathway Maps  
**Authors**: Kaibo Zhang, Ji Wu, Chao Zhang, Andrew Thwaites  
**arXiv**: 2607.06284v1 (2026-07-07)  
**Categories**: q-bio.NC, stat.AP

## Core Methodology

### Problem Statement

Information Processing Pathway Maps (IPPMs) formalize the sequence of mathematical transformations applied to sensory stimuli, charting latency and cortical expression of computational steps. The challenge: **how to determine which of several competing computational models best explains neural data?**

Traditional approach: Frequentist hypothesis testing (reject null hypothesis)  
Proposed approach: Bayesian probabilistic inference (quantify relative evidence)

### Key Innovation: Paradigm Shift

**From**: "Is this model significantly different from null?"  
**To**: "How much evidence supports model A vs model B vs model C?"

This shift from **hypothesis rejection** to **model adjudication** better captures the scientific question: we're not testing if a model is "significant" - we're comparing which computational theory best explains the data.

### Information Processing Pathway Maps (IPPMs)

**Definition**: Scalable framework for formalizing sensory processing as a sequence of transformations:
1. Stimulus → Computational step 1 (latency τ₁, cortical region R₁)
2. → Computational step 2 (latency τ₂, cortical region R₂)
3. → ... → Predicted neural response

**Core Components**:
- **Computational models**: Mathematical transformations (e.g., spectrogram, modulation spectrum)
- **Temporal parameters**: Latency of each processing stage
- **Spatial parameters**: Cortical regions expressing each computation
- **Statistical mapping**: Link model outputs to observed neural activity

### Frequentist vs Bayesian Approaches

#### Frequentist Approach (Traditional)

```
For each computational model M:
  H₀: Model M does not explain neural data
  H₁: Model M explains neural data
  
  Compute test statistic (e.g., correlation, F-statistic)
  Calculate p-value = P(data | H₀)
  If p < α (e.g., 0.05): reject H₀, conclude model is "significant"
```

**Limitations**:
- Binary decision (significant/not) loses information
- Cannot compare models directly (each tested against null separately)
- p-value ≠ probability that model is correct
- Collinear models (highly correlated predictions) cause multiple testing issues

#### Bayesian Approach (Proposed)

```
Define competing models: {M₁, M₂, ..., Mₖ}
Compute posterior probability: P(Mᵢ | data) ∝ P(data | Mᵢ) × P(Mᵢ)

Compare models via:
- Bayes Factor: BF₁₂ = P(data | M₁) / P(data | M₂)
- Posterior odds: P(M₁ | data) / P(M₂ | data)
- Model probabilities: P(Mᵢ | data) for all i
```

**Advantages**:
- Direct model comparison (not vs null)
- Quantifies evidence strength (not just binary decision)
- Handles collinear models naturally
- Accumulates evidence across studies
- Incorporates prior knowledge

### Implementation Framework

#### Step 1: Define Computational Models

Specify candidate transformations for sensory processing:
```python
# Example: Auditory processing pathway
models = {
    'spectrogram': compute_spectrogram,      # Time-frequency representation
    'modulation_spectrum': compute_modspec,  # Temporal modulation analysis
    'envelope': compute_envelope,            # Amplitude envelope
    'edge_detection': compute_spectral_edges # Spectral contrast features
}
```

#### Step 2: Generate Predictions

For each model, predict neural response at each latency and cortical location:
```python
for model_name, model_func in models.items():
    for latency in range(0, 300, 10):  # 0-300ms in 10ms steps
        predicted_response = model_func(stimulus, latency)
        # Compare to observed neural data (e.g., EEG, MEG, fMRI)
```

#### Step 3: Compute Evidence

**Frequentist**:
```python
from scipy import stats

for model_name, predictions in model_outputs.items():
    correlation = np.corrcoef(predictions, observed_data)[0, 1]
    p_value = stats.ttest_1samp(correlations, 0).pvalue
    # Binary: significant if p < 0.05
```

**Bayesian**:
```python
import pymc3 as pm

with pm.Model() as model_comparison:
    # Define likelihood for each computational model
    for model_name, predictions in model_outputs.items():
        sigma = pm.HalfNormal(f'sigma_{model_name}', sigma=1.0)
        pm.Normal(f'obs_{model_name}', mu=predictions, sigma=sigma, 
                  observed=observed_data)
    
    # Compute Bayes factors via marginal likelihoods
    trace = pm.sample(1000, tune=1000)
    # Use bridge sampling or importance sampling for marginal likelihood
```

#### Step 4: Reconstruct Pathway

Select model with highest posterior probability at each latency/region:
```python
pathway_map = {}
for latency in latencies:
    for region in cortical_regions:
        # Compare all models at this spatiotemporal location
        posterior_probs = compute_posteriors(models, data, latency, region)
        best_model = max(posterior_probs, key=posterior_probs.get)
        pathway_map[(latency, region)] = {
            'model': best_model,
            'probability': posterior_probs[best_model],
            'evidence_strength': interpret_bayes_factor(posterior_probs)
        }
```

### Case Study: Auditory Loudness Processing

**Dataset**: Auditory neuroimaging (MEG/EEG) with sound stimuli  
**Known pathway**: Cochlear nucleus → Inferior colliculus → Auditory cortex (loudness processing)

**Models compared**:
1. Linear envelope (amplitude)
2. Log-compressed envelope (perceived loudness)
3. Spectral contrast (frequency-specific energy)
4. Modulation spectrum (temporal dynamics)

**Results**:
- Bayesian approach correctly identified log-compressed envelope as dominant in auditory cortex
- Frequentist approach showed multiple "significant" models (due to collinearity)
- Bayesian provided graded evidence: log-envelope (70% prob) > linear (20%) > others (10%)

## Practical Applications

### Use Cases

1. **Sensory Processing Research**: Map computational steps in vision, audition, somatosensation
2. **Model Comparison**: Compare competing theories of neural computation
3. **Clinical Applications**: Identify disrupted processing pathways in neurological disorders
4. **Brain-Computer Interfaces**: Select optimal feature extraction pipeline

### When to Use Bayesian vs Frequentist

**Use Bayesian when**:
- Comparing multiple competing models (not just vs null)
- Need to quantify evidence strength (not just significance)
- Models are collinear (highly correlated predictions)
- Accumulating evidence across studies
- Prior knowledge available

**Use Frequentist when**:
- Simple hypothesis testing (model vs null)
- Large sample sizes (asymptotic assumptions hold)
- Computational resources limited (Bayesian methods slower)
- Standard practice in field (for comparability)

## Key Insights

1. **Model Adjudication > Hypothesis Testing**: The scientific question is "which model is best?" not "is this model significant?" - Bayesian approach directly answers the right question
2. **Collinearity Handling**: Bayesian methods naturally handle correlated model predictions without multiple testing corrections
3. **Evidence Accumulation**: Bayes factors can be combined across studies (multiplicative), enabling meta-analysis
4. **Interpretability**: Posterior probabilities are intuitive (P(model | data)), unlike p-values (P(data | null))

## Limitations & Considerations

**Computational Cost**: Bayesian methods require MCMC sampling or numerical integration (slower than frequentist)

**Prior Specification**: Results depend on prior choices - sensitivity analysis required

**Model Space**: Must define candidate models upfront - cannot discover new models

**Implementation Complexity**: Requires Bayesian statistical expertise (PyMC3, Stan, etc.)

## Implementation Resources

**Python Libraries**:
- PyMC3/PyMC4: Bayesian modeling with MCMC
- Stan (via CmdStanPy): High-performance Bayesian inference
- BayesFactor: R package for Bayes factor computation

**Tutorials**:
- "Statistical Rethinking" by Richard McElreath (Bayesian modeling)
- "Doing Bayesian Data Analysis" by John Kruschke

## Activation Triggers

Use this skill when:
- Mapping sensory processing pathways (auditory, visual, somatosensory)
- Comparing competing computational models of neural data
- Quantifying evidence for neural entrainment
- Building Information Processing Pathway Maps
- Choosing between frequentist and Bayesian approaches for model selection

## Related Concepts

- **Neural Entrainment**: Phase-locking of neural oscillations to stimulus features
- **Temporal Response Function (TRF)**: Linear mapping from stimulus to neural response
- **Model Selection**: AIC, BIC, cross-validation (frequentist alternatives)
- **Bayesian Model Averaging**: Weight predictions by model posterior probabilities