predictive-subspace-recovery-profiles

name

predictive-subspace-recovery-profiles

description

Target-Space Recovery Profiles methodology for evaluating model-brain alignment beyond prediction accuracy. Identifies which reproducible brain response dimensions are recovered by predictions, enabling diagnostic evaluation of NeuroAI model-brain alignment. Activation: model brain alignment, predictive subspace, recovery profile, brain prediction evaluation, NeuroAI alignment, target-space recovery, NSD analysis.

Predictive Subspace Recovery Profiles Methodology

Target-Space Recovery Profiles methodology for evaluating model–brain alignment beyond scalar prediction accuracy.

Core Insight

Prediction accuracy alone cannot reveal which dimensions of the target brain's response space are recovered by a model. Two models with identical prediction accuracy may recover completely different response dimensions. This methodology makes the structural content of prediction explicit.

Key Components

1. Reproducible Target Reference

• Use repeated fMRI measurements of the same stimuli to identify target-brain response dimensions that are reproducibly recoverable across independent trial splits
• Fit target-to-target predictions between split halves → extract orthonormal target basis vectors uⱼ ranked by reproducibility
• The first 3 dimensions typically account for ~89% of normalized reference weight; median entropy effective rank ≈ 5.12

2. Predictive Subspace

• For any source (another subject's brain OR a model's internal representations), fit a ridge-regularized low-rank linear mapping to target responses
• Extract the orthonormal basis Qₛ spanning the predictive subspace in target response space
• Prediction accuracy = held-out correlation between predicted and observed responses

3. Recovery Profile

• Directional Reference Coverage: DirCovₛ,ⱼ = ‖Qₛᵀuⱼ‖²₂ — how much the predictive subspace overlaps with each target-reference direction
• Sort target-reference dimensions by coverage strength → get ordered profile
• Profile shape reveals which dimensions are recovered and which are missed

4. Brain-to-Brain Human Reference

• Brain-to-brain recovery profile provides human reference: shows which dimensions are typically recoverable from another subject
• Model-to-brain profiles should be interpreted relative to this human reference
• Coverage declines from ~0.96 at k=1 to ~0.87 at k=10 for brain-to-brain (structured decline, not flat)

Key Findings (Nakamura et al., 2026)

Pretraining changes recovery profiles beyond accuracy

• ImageNet-pretrained models exceed 4-seed random mean in profile mean by 0.177 (95% CI: 0.162–0.192)
• Random models can match prediction accuracy but recover different dimensions

Brain-to-brain as diagnostic reference

• Brain sources recover dimensions with characteristic declining profile
• Provides "human ceiling" for what is recoverable from biological systems
• Pairs of brain sources show small profile differences; model-brain pairs show larger differences

Accuracy-matched analysis reveals structural mismatches

• Near-equal accuracy pairs (|Δaccuracy| ≤ 0.01): pretrained models consistently show higher top-k coverage
• Scalar accuracy masks directional mismatches in target response space

When to Use

• Requires repeated measurements of target brain (splits define reference)
• Most appropriate for measured response patterns (voxel, neural population responses)
• Complementary to existing alignment methods (RSA, encoding models, alignment pattern analysis)
• Useful when goal is not only to predict but to understand which parts of target response space are recovered

Experimental Protocol

Data

• NSD-core-shared: 8 subjects, 515 shared repeated natural images
• ROIs: V1v, V1d, V2v, V2d, V3v, V3d, hV4 (both hemispheres)
• 5 synchronized outer folds for cross-validation

Sources

• Brain sources: responses from non-target subjects in corresponding ROI
• Model sources: ResNet-18/50, VGG-16, ViT-B/16 (pretrained and randomly initialized)

Fits

• Ridge-regularized low-rank linear fits
• Rank and regularization selected by inner CV on outer-training images
• Recovery profiles computed from source-induced predictive subspaces
• Profile plots display top-k prefixes through k=10

Implementation Steps

1. Estimate reproducible target reference via split-half target-to-target prediction
2. Fit source-to-target mapping for each source (brain or model)
3. Extract predictive subspace Qₛ from fitted mapping
4. Compute directional coverage for each target-reference direction
5. Build recovery profile by sorting dimensions by coverage strength
6. Compare model-to-brain profile against brain-to-brain human reference

Pitfalls

• Not a global brain-likeness claim: high recovery only supports conditional interpretation for the evaluated ROI, dataset, preprocessing, and readout class
• Cannot be used as held-out prediction accuracy: recovery profiles are diagnostics, not additional prediction scores
• Repeated measurements required: without repeats, reproducible reference cannot be estimated
• Fitting/evaluation separation: outer-test responses only define evaluation reference, never used for model selection or hyperparameter tuning

Related Concepts

• Encoding models, Representational Similarity Analysis (RSA)
• Alignment Pattern Analysis (cross-region relational criterion)
• Spectral theory of neural prediction (model-side geometry decomposition)
• Brain-to-brain prediction as human reference benchmark
• GLMdenoise response-amplitude estimation

Related Skills

• decoding-encoding-alignment-critique — Critical analysis of brain-model alignment methods
• feature-visualization-brain-encoder — Feature visualization for brain encoder interpretability
• naturality-violation-score — Category-theoretic brain-DNN alignment
• brain-dit-fmri-foundation-model — fMRI foundation model evaluation
• in-context-brain-decoding — Training-free cross-subject brain decoding

Reference

• Title: Beyond Prediction Accuracy: Target-Space Recovery Profiles for Evaluating Model–Brain Alignment
• Authors: Ken Nakamura, Tomoya Nakai, Ryuto Yashiro, Ayumu Yamashita, Kaoru Amano
• arXiv: 2605.20127 [q-bio.NC, cs.AI, cs.LG]
• Date: May 2026
• Institution: The University of Tokyo, Osnabrück University, Freie Universität Berlin, Kobe University
• Dataset: Natural Scenes Dataset (NSD)
• URL: https://arxiv.org/abs/2605.20127