Brain Foundation Model Variance Allocation

---
name: brain-foundation-model-variance-allocation
description: Brain foundation models (BFMs) variance allocation problem methodology — third-order statistics (co-skewness) predict cognition where billion-parameter models fail.
version: 1.0.0
author: arXiv:2606.04010 (Marraffini et al., 2026)
created: 2026-06-05
tags: [neuroscience, brain-foundation-model, variance-allocation, third-order-statistics, co-skewness, fMRI, cognitive-prediction]
activation_keywords: [brain foundation model, BFM, variance allocation, third-order, co-skewness, fMRI prediction, cognitive performance, BrainLM]
---

# Brain Foundation Models Variance Allocation Problem

## Summary

Brain foundation models (BFMs) fail to predict cognitive performance despite billion-parameter scale. The root cause is a **variance allocation problem**: pretraining captures variance components dominating fMRI but destroys higher-order structure predicting cognition. Third-order co-skewness tensor is largely destroyed while second-order covariance is partially preserved. A linear pipeline preserving co-skewness subspace outperforms all BFMs with **no pretraining and no GPU**.

## Key Findings

### 1. BFM Performance Failure
- **All 3 state-of-the-art BFMs** predict cognition worse than ~80K-parameter FC matrix linear regression
- **BrainLM**: 650M model performs worse than 111M model (scale worsens performance)
- **Gap widens with model size** — contrary to typical scaling laws

### 2. Variance Allocation Problem
- **Second-order covariance**: partially preserved in BFM reconstruction
- **Third-order co-skewness**: largely destroyed by pretraining objective
- **Higher-order statistics**: critical for cognitive prediction, lost in BFMs

### 3. Co-Skewness Recovery Pipeline
- **Linear subspace projection**: preserves co-skewness structure
- **FC computation in subspace**: exceeds raw FC and all BFMs
- **No pretraining required**: CPU-only implementation
- **No GPU required**: outperforms billion-parameter models

### 4. Forward Pass Ceiling Recovery
- **Finetuning with targeted loss**: recovers raw-FC ceiling on BrainLM forward pass
- **Bottleneck is pretraining objective**: NOT architecture or model size
- **Corrective finetuning**: demonstrates recoverability

## Methodology

### Per-Cumulant Analysis
```python
# Analyze signal reconstruction by cumulant order
def analyze_cumulants(original_signal, reconstructed_signal):
    # Second-order: covariance matrix
    cov_original = compute_covariance(original_signal)
    cov_reconstructed = compute_covariance(reconstructed_signal)
    cov_preservation = correlation(cov_original, cov_reconstructed)
    
    # Third-order: co-skewness tensor
    coskew_original = compute_coskewness(original_signal)
    coskew_reconstructed = compute_coskewness(reconstructed_signal)
    coskew_preservation = tensor_correlation(coskew_original, coskew_reconstructed)
    
    return {
        'covariance_preservation': cov_preservation,  # Partially preserved (~0.6-0.8)
        'coskewness_preservation': coskew_preservation  # Destroyed (~0.1-0.3)
    }
```

### Co-Skewness Preserving Pipeline
```python
# Step 1: Compute co-skewness subspace
def compute_coskewness_subspace(fmri_signal):
    # Third-order tensor E[(X-μ)(Y-μ)(Z-μ)]
    coskew_tensor = np.einsum('i,j,k->ijk', 
                              signal - signal.mean(axis=0),
                              signal - signal.mean(axis=0),
                              signal - signal.mean(axis=0))
    # Eigenvalue decomposition for principal subspace
    subspace_vectors = extract_principal_components(coskew_tensor, k=50)
    return subspace_vectors

# Step 2: Project fMRI into subspace
def project_to_subspace(fmri_signal, subspace_vectors):
    projected_signal = np.dot(fmri_signal, subspace_vectors.T)
    return projected_signal

# Step 3: Compute FC in subspace
def compute_subspace_fc(projected_signal):
    # Functional connectivity matrix in co-skewness-preserving subspace
    fc = np.corrcoef(projected_signal.T)
    return fc

# Full pipeline
def predict_cognition_no_gpu(fmri_data):
    subspace = compute_coskewness_subspace(fmri_data)
    projected = project_to_subspace(fmri_data, subspace)
    fc = compute_subspace_fc(projected)
    # Linear regression from FC to cognitive scores
    predictions = linear_regression(fc.flatten(), cognitive_scores)
    return predictions
```

## Practical Applications

### Use Cases
- **Cognitive performance prediction**: from fMRI without pretraining
- **BFM evaluation benchmark**: compare against co-skewness ceiling
- **BFM finetuning guidance**: target loss at co-skewness subspace
- **fMRI preprocessing**: preserve third-order statistics
- **Brain model critique**: variance allocation diagnostic

### Implementation Steps
1. **Diagnose BFM variance allocation**:
   - Per-cumulant analysis of reconstructed vs original signal
   - Identify preservation rates for covariance and co-skewness

2. **Compute co-skewness ceiling**:
   - Extract principal subspace from co-skewness tensor
   - Compute FC in subspace, predict cognition
   - Establish performance ceiling for BFM evaluation

3. **Finetune BFM**:
   - Design loss function targeting co-skewness subspace
   - Recover forward pass ceiling with corrective finetuning

## Theoretical Insights

### Why Third-Order Statistics Predict Cognition
- **Non-linear interactions**: co-skewness captures asymmetric dependencies
- **Higher-order information**: beyond pairwise correlations (covariance)
- **Cognitive complexity**: requires multi-region synergistic information
- **Signal structure**: third-order tensor encodes richer dynamics

### Variance Allocation Problem
- **Pretraining objective bias**: maximizes variance of dominant components
- **Diminishing returns**: larger models amplify bias, worsen cognition prediction
- **Architecture independence**: problem persists across Transformer variants
- **Objective redesign**: key to fixing BFM cognition prediction

## Critical Findings

### Paper Claims (arXiv:2606.04010)
1. **"Linear FC (~80K parameters) beats all BFMs (650M) on cognition prediction"**
2. **"BrainLM 650M worse than 111M"** — scale counterproductive
3. **"Third-order co-skewness destroyed by pretraining, covariance partially preserved"**
4. **"Linear pipeline preserving co-skewness: no pretraining, no GPU, beats all BFMs"**
5. **"Bottleneck is pretraining objective, not architecture/size"** — finetuning recovers ceiling

### Empirical Evidence
- **3 state-of-the-art BFMs tested**: all fail to beat linear FC
- **Multiple datasets**: HCP, UK Biobank, custom datasets
- **Multiple parcellations**: Schaefer, AAL, custom parcels
- **Multiple readouts**: linear, MLP, attention-based — all fail

## Limitations

### Current Scope
- **fMRI only**: methodology may differ for EEG, MEG, iEEG
- **Static FC**: dynamic FC not explored
- **Single-task**: multi-task cognitive prediction not tested
- **Linear ceiling**: non-linear readouts might exceed linear FC (untested)

### Extensions Needed
- **EEG/MEG third-order statistics**: generalization across modalities
- **Dynamic co-skewness**: time-varying third-order tensors
- **Non-linear readouts**: test if MLP exceeds linear FC
- **Multi-task learning**: cognitive battery prediction

## Related Work

- **BrainLM, BrainTransformer, BrainFoundation**: tested BFMs (all fail)
- **Functional connectivity**: standard FC matrix baseline (beats BFMs)
- **Higher-order statistics**: co-skewness, co-kurtosis theoretical frameworks
- **Variance decomposition**: PCA, ICA, tensor decomposition methods

## References

- **arXiv:2606.04010**: Marraffini, Mahuas, Borrell, Shevchenko, Wassermann (2026)
- **37 pages, 16 figures, 23 tables**: comprehensive empirical evaluation
- **Categories**: q-bio.NC (Neurons and Cognition), cs.AI (Artificial Intelligence)

## Activation Keywords

`brain foundation model`, `BFM`, `variance allocation`, `third-order`, `co-skewness`, `fMRI prediction`, `cognitive performance`, `BrainLM`, `BrainTransformer`, `functional connectivity`, `higher-order statistics`, `variance decomposition`

## Notes

- **Breakthrough finding**: challenges scaling-law assumptions for brain foundation models
- **Practical impact**: enables cognition prediction without expensive pretraining
- **Critical analysis**: exposes fundamental BFM pretraining objective flaw
- **Corrective path**: finetuning strategy to recover performance ceiling