Neural Co-Evolution Visualizer

Comparing representational reorganisation in biological and artificial neural systems during concept acquisition
DATASET COMPLETE -- April 2026
Methodological note (per Hillis, April 2026): Participants in the scanner are not passively viewing photographs. They are performing the Free Body Diagram (FBD) task -- actively identifying and evaluating Newtonian forces acting on each structure. The dissimilarity matrices therefore encode physics reasoning patterns, not visual processing. See Cetron et al. (2019) Nature Communications 10:2027 for FBD task protocol.
Dataset Status
Subjects
50
26 lab / 25 textbook
Brain Parcels
300
Schaefer parcellation
Stimuli
12
3 categories x 4 items
Data Points
990,000
Neural distances
Representational Geometry

Biological Neural DM

50-subject average | FBD task activations
Cantilever Truss Vertical Load

ANN Representational DM

Transformer hidden states | Same 12 stimuli
Correlation distance from [CLS] token representations
Representational Convergence
Spearman r between biological and ANN dissimilarity matrices
--
Behavioural Data (uploaded April 23, 2026)
FBD Task Accuracy
--
Mean across 50 subjects (range 0-85)
Quiz 1 Mean
--
Topic 1 concept assessment
Quiz 2 Mean
--
Topic 2 concept assessment
Subject Behavioural Scores
Subject Group FBD Quiz 1 Quiz 2 Pooled
Data Provenance
EXACT 50 subjects x 300 parcels x 66 neural distances (7.6 MB pickle)
EXACT 12x12 category model matrix and stimulus legend
EXACT Complete analysis pipeline: Scripts 01 (Group SVM), 02 (Neural Scores), 03 (LMER)
NEW Behavioural scores: FBD accuracy, Quiz 1, Quiz 2, QuizPooled (50 subjects)
EXACT Teaching materials: Statics Primer, Topic 1-2 slides, lab sheets, inquiry sheets
EXACT 160 significant parcels identified via group-level SVM (Script 01)
The COEVO Hypothesis

Biological Pathway

Human performs FBD task on 12 structures
Human receives physics instruction (statics + topics)
fMRI measures neural DMs at 300 parcels
SVM classifies categories from DMs
Neural score correlates with quiz performance

ANN Pathway (COEVO Simulation)

Transformer processes same 12 structure descriptions
Transformer fine-tuned on same physics concepts (LoRA)
Hidden-layer activations produce DMs at each layer
Same SVM pipeline classifies categories from ANN DMs
ANN neural score compared against human neural scores
Evidence Chain Assessment
STRONG Link 1-2: Learning produces representational reorganisation (Cetron 2026) + BDNF markers (Jinich-Diamant 2025)
MODERATE Link 3: BDNF promotes neurogenesis in adult hippocampus (Dumitru et al. 2025, Science)
WEAKEST Link 4: Reorganisation = neurogenesis (correlated but not proven causally in single study)
TESTABLE Link 5: ANN shows functional equivalent of reorganisation (THIS is what the simulation tests)