| name | feature-visualization-brain-encoder |
| description | Feature visualization as interpretability technique for brain encoder models. Uses gradient ascent on predicted activation for target ROIs to qualitatively evaluate whether encoders have internalized functional brain organization. Activation: feature visualization brain encoder, cortical selectivity validation, brain encoder interpretability, ROI feature visualization. |
Feature Visualization for Brain Encoder Interpretability
Proposes feature visualization as a complementary interpretability technique for brain encoder models, going beyond held-out prediction accuracy to assess whether encoders have internalized the functional organization of the brain.
Metadata
- Source: arXiv:2605.13904
- Authors: Stuart Bladon, Brinnae Bent
- Published: 2026-05-13
- Paper: "Feature Visualization Recovers Known Cortical Selectivity from TRIBE v2"
Core Methodology
Problem
Brain encoder models are typically evaluated by held-out prediction accuracy -- useful for training but poor for interpretation. High prediction scores don't reveal whether the model has internalized the brain's functional organization.
Key Innovation
Feature Visualization for Brain Encoders: Gradient ascent on the encoder's predicted activation for a target region of interest (ROI), synthesizing images that maximally activate each brain region.
Technical Framework
- Compose brain encoder with differentiable backbone (e.g., TRIBE v2 + V-JEPA 2 ViT-G)
- Hold both frozen during optimization
- Gradient ascent on input image pixels to maximize predicted activation for target ROI
- Evaluate synthesized images against known cortical selectivity patterns
Results
- V1-V4 progression: Recovered visible hierarchy of increasing spatial scale and feature complexity
- MT (middle temporal): Radial "frozen-motion" streaks despite static-only optimization
- FFA (fusiform face area): Face-like features, optimized stimuli drive ~4x more than natural faces (adversarial super-stimuli)
- PPA (parahippocampal place area): Consistent rectilinear line patterns
Applications
- Qualitative evaluation of brain encoders
- Validation of in-silico neuroscience models
- Discovery of unexpected selectivity patterns
- Applicable to any brain encoder with differentiable backbone
Pitfalls
- Optimized stimuli are super-stimuli, not canonical exemplars
- Requires differentiable encoder backbone
- Static optimization may miss temporal selectivity (e.g., MT's motion preference only partially recovered)
Related Skills
- tribe-v2-foundation-model
- tribe-v2-trimodal-foundation-model
- decoding-encoding-alignment-critique
- lpact-brain-lm-alignment-evaluation
