| name | brain-dnn-transformation-alignment |
| description | Naturality Violation Score (NVS) for brain-DNN alignment beyond object-level comparison. Uses category theory to test whether brains and DNNs preserve the same transformations among stimuli. Covers approximate naturality, axis-resolved alignment analysis, and hierarchy crossover detection. Use when: (1) evaluating brain-DNN alignment at the transformation level, (2) comparing neural representations across model architectures, (3) analyzing semantic vs low-level visual alignment, or (4) designing neuroscience-inspired model evaluation. Activation: naturality violation score, NVS, brain-DNN alignment, transformation alignment, category theory neuroscience, axis-resolved alignment, hierarchy crossover, approximate naturality, brain model comparison.
|
Brain-DNN Transformation Alignment (NVS)
Based on Kamitani (2026) — arXiv:2605.06420.
Core Concept
Traditional brain-DNN alignment measures per-stimulus correspondence (encoding
accuracy, RSA, CKA). This paper introduces a transformation-level alignment
test: do the brain and model preserve the same transformations among stimuli?
Mathematical Framework
Approximate Naturality
Inspired by category theory, alignment is formalized as naturality:
Brain side: x ──f──► x'
│ │
φ_brain φ_model
↓ ↓
Model side: φ(x)─f'──► φ(x')
If the brain and model preserve the same transformation f, the naturality
square approximately commutes:
φ_model(f(x)) ≈ f'(φ_brain(x))
Naturality Violation Score (NVS)
NVS quantifies deviation from commutativity, normalized to a permutation null:
NVS(f) = ||φ_model(f(x)) - f'(φ_brain(x))|| / E_perm[||...||]
- NVS ≈ 0: perfect transformation preservation
- NVS ≈ 1: no better than random permutation
- Lower NVS = better alignment for that transformation axis
Key Findings
Hierarchy Crossover
Applied to fMRI (GOD dataset, 5 subjects) + 3 vision DNNs + 3 world-model
proxies:
| Alignment Axis | Best Aligned To | NVS |
|---|
| Semantic (animacy) | High-level visual cortex + deep DNN layers | 0.39 |
| Mid-level visual | Mid visual cortex + mid DNN layers | < next-best |
| Low-level visual | Early visual cortex + shallow layers | < next-best |
This reveals that alignment is axis-selective rather than uniform —
different transformation families align at different depths.
Dissociation from Traditional Metrics
- NVS captures alignment failures that RSA/CKA cannot resolve
- Encoding/decoding accuracy does not predict NVS
- Alignment is selective over candidate morphism families
Implementation Workflow
Step 1: Define Proxy Space
Choose a proxy space W and comparison maps:
- Proxy: semantic embeddings (e.g., CLIP, Word2Vec)
- Maps: morphisms between stimuli (e.g., animacy gradient, pose rotation)
Step 2: Compute NVS Per Axis
For each transformation axis f:
- Apply f to stimuli on brain side → get brain responses
- Apply f to stimuli on model side → get model activations
- Measure commutativity deviation
- Normalize against permutation null distribution
Step 3: Axis-Resolved Analysis
- Compare NVS across multiple axes (semantic, spatial, temporal)
- Identify which brain regions align with which model layers per axis
- Detect hierarchy crossovers (different axes peak at different depths)
Practical Applications
Model Selection
Use NVS to choose DNN architectures that best preserve brain-relevant
transformations for a specific task domain.
Layer-wise Alignment
Map which DNN layers correspond to which brain areas for different
transformation types — useful for neuroscientific interpretability.
Proxy Space Design
The method generalizes to any proxy space (world models, symbolic
representations, etc.), enabling richer brain-model comparisons.
Comparison with Existing Methods
| Method | Granularity | Transformation-aware? | Null model |
|---|
| Encoding accuracy | Per-stimulus | No | None |
| RSA | Geometry | Partial | None |
| CKA | Global similarity | No | None |
| NVS (this work) | Per-axis | Yes | Permutation null |
Common Pitfalls
- Proxy space quality: NVS results depend on proxy choice; poor proxies
yield misleading NVS values
- Transformation definition: candidate morphisms must be well-defined
and meaningful for both brain and model domains
- Permutation null: ensure sufficient permutations (≥1000) for stable
normalization
- W-less control: always run anchor-ablation to verify alignment isn't
driven by trivial correlations
Activation Keywords
- naturality violation score, NVS, brain-DNN transformation alignment,
approximate naturality, axis-resolved alignment, hierarchy crossover,
category theory neuroscience, brain model comparison, transformation
preservation, semantic alignment