decoding-encoding-alignment-critique

---
name: decoding-encoding-alignment-critique
description: >
  Critical analysis framework for brain-model alignment methods (RSA, encoding, decoding).
  Exposes that decoding alignment does NOT imply computational similarity — high RSA/DSA
  scores can arise from small non-representative neuron subpopulations. Introduces encoding
  manifolds as complementary analysis tool. Covers representational collapse, feature
  confounding, stimulus-set dependency, and subpopulation dominance.
  Use when: (1) evaluating brain-model alignment methodologies, (2) designing RSA studies,
  (3) interpreting encoding/decoding results, (4) selecting appropriate alignment metrics,
  (5) reviewing neuroscience-AI alignment papers, (6) avoiding alignment pitfalls.
  Trigger words: RSA critique, encoding model limitations, representational similarity pitfalls,
  brain-model alignment, representational collapse, feature confounding, alignment validity,
  stimulus-set dependency, cross-decoding, transformation alignment, encoding manifold,
  subpopulation dominance, neural system comparison.
---

# Decoding-Encoding-Alignment Critique

Critical framework for evaluating and interpreting brain-model alignment methods.

## Core Thesis

Standard alignment metrics (RSA, encoding models, linear decoding) have systematic blind spots
that can produce misleading conclusions about brain-AI similarity. **Decoding alignment does NOT
imply computational similarity.**

## Key Limitations Exposed

### 1. Subpopulation Dominance (NEW - arXiv:2605.05907)

**Problem**: High RSA/DSA alignment can arise from tiny, non-representative subpopulations of
neurons. The representational geometry reflects a few neurons, not the whole population.

**Evidence**:
- Similar decoding behavior and high alignment from small subpopulations while overall encoding
  topology differs completely
- Causal MNIST experiment: decoding metrics unchanged when encoding topology is manipulated
  via training loss
- Alignment metrics are blind to how function is distributed across neurons

**Detection**:
```python
def subpopulation_rsa(neural_responses_a, neural_responses_b, n_range=[5,10,20,50,100]):
    """Test RSA stability across random subpopulations."""
    import numpy as np
    from scipy.spatial.distance import pdist, squareform
    from scipy.stats import spearmanr
    
    results = {}
    for n in n_range:
        if n >= min(neural_responses_a.shape[0], neural_responses_b.shape[0]):
            continue
        rsa_scores = []
        for _ in range(100):
            idx_a = np.random.choice(neural_responses_a.shape[0], n, replace=False)
            idx_b = np.random.choice(neural_responses_b.shape[0], n, replace=False)
            rdm_a = squareform(pdist(neural_responses_a[idx_a], metric='correlation'))
            rdm_b = squareform(pdist(neural_responses_b[idx_b], metric='correlation'))
            rsa_scores.append(spearmanr(rdm_a.ravel(), rdm_b.ravel())[0])
        results[n] = {'mean': np.mean(rsa_scores), 'std': np.std(rsa_scores)}
    return results
```

### 2. Encoding-Decoding Decoupling (NEW - arXiv:2605.05907)

**Problem**: Alignment metrics capture WHAT is represented (decoding) but not HOW it's
implemented (encoding). Two systems can have identical decoding but completely different
encoding structures.

**Solution - Dual Manifold Analysis**:
| Aspect | Decoding Manifold | Encoding Manifold |
|--------|-------------------|-------------------|
| What it measures | Stimulus distinguishability from neural activity | How neurons are organized across responses |
| Unit of analysis | Stimulus space geometry | Neural population space geometry |
| Answers | "What information is represented?" | "How is it implemented across neurons?" |
| Sensitivity | Insensitive to which neurons contribute | Captures global neuronal organization |

```python
# Encoding RDM: pairwise similarity of neuron tuning curves
def compute_encoding_rdm(neural_responses):
    """Rows are neurons; captures neuron-neuron similarity structure."""
    from scipy.spatial.distance import pdist, squareform
    return squareform(pdist(neural_responses.T, metric='correlation'))
```

### 3. Representational Collapse

**Problem**: RSA can report high similarity when both representations are degenerate/collapsed.
Two uninformative representations can have high RSM correlation.

**Detection**: Check representational dimensionality (participation ratio, intrinsic dimension).
Compare against low-dimensional baselines.

### 4. Feature Confounding

**Problem**: Encoding models can achieve high performance by capturing low-level confounds
(pixel statistics, image size, contrast) rather than semantic representations.

**Detection**: Control for low-level features; cross-validate across stimulus sets.

### 5. Stimulus-Set Dependency

**Problem**: Alignment scores vary dramatically with stimulus selection.

**Detection**: Test on multiple diverse stimulus sets; report cross-stimulus generalization.

### 6. Linear Probing Artifacts

**Problem**: High linear decoding accuracy only proves information is linearly extractable,
not that the model uses the same representations as the brain.

## Recommended Evaluation Protocol

1. **Dual-manifold analysis**: Always pair decoding (RSA/DSA) with encoding analysis
2. **Subpopulation sensitivity**: Test alignment stability across neuron subsamples
3. **Dimensionality matching**: Compare representations at matched dimensionalities
4. **Cross-stimulus validation**: Test generalization across stimulus distributions
5. **Null model comparison**: Include trivial baselines
6. **Transformation analysis**: Test if representations transform similarly, not just match statically

## Common Pitfalls in Alignment Studies

| Pitfall | Symptom | Fix |
|---------|---------|-----|
| Subpopulation dominance | High RSA driven by few neurons | Subpopulation stability analysis |
| Encoding blindness | Identical decoding, different implementation | Encoding manifold analysis |
| RSA inflation | High RSM correlation with degenerate RSMs | Check RSM condition number |
| Encoding overfitting | High R² but poor generalization | Cross-validate across stimulus sets |
| Decoding ceiling | Near-perfect decoding from both | Use matched-capacity probes |
| Category confound | Alignment driven by coarse categories | Control for category structure |

## When to Use

- Evaluating brain-DNN alignment beyond RSA
- Comparing neural coding across brain regions
- Understanding how information is distributed in populations
- Critiquing representational similarity studies
- Designing comprehensive neural system comparisons

## arXiv References

### Primary: Subpopulation Dominance in RSA
- Paper: "Decoding Alignment without Encoding Alignment: A critique of similarity analysis in neuroscience"
- Authors: Bertram, J., Dyballa, L., Keller, T.A., Kinger, S., & Zucker, S.W. (2026)
- ID: arXiv:2605.05907v1 | Category: q-bio.NC | Date: 2026-05-07
- Key finding: RSA alignment can be driven by tiny subpopulations; encoding topology must be analyzed complementarily

### Secondary: Transformation Alignment
- Paper: "Beyond Object-Level Alignment: Do Brains and DNNs Preserve the Same Transformations?"
- Author: Yukiyasu Kamitani (2026)
- ID: arXiv:2605.06420v1 | Category: q-bio.NC
- Proposes category-theoretic approach: do brain and model preserve the same candidate transformations?