Cortex v1 · Research preview

Brain pattern recognition for neuroscience research

Cortex reads resting-state brain connectivity and identifies patterns associated with autism — a 227M-parameter transformer trained on 1,545 real fMRI subjects. Three research tools. No clinical claims.

227M
parameters
1,545
subjects
63%
val accuracy

Forward, and inverse

AQAL predicts how a neurodiverse brain experiences a stimulus. Cortex runs the other direction — it reads an existing scan and characterizes the neural pattern.

Two complementary halves of the same pipeline: one maps stimulus to brain, the other maps brain to classification.

AQAL
Stimulus → Brain

Video, audio, or text in. Predicts neurodiverse response. For spaces, therapy, sensory audits.

Cortex
Brain → Classification

Connectivity in. Identifies ASD-associated patterns. For stratification, QC, biomarkers.

Cortex Architecture

A 227M-parameter transformer over Fisher-z connectivity (100-ROI Schaefer atlas), pooled into a subject embedding and branched into three joint heads.

Connectivity(B, 4950)
Projection32 × 1024
16× Encoder16 heads
Latent256-dim
Classifier → ASD
Reconstructor
01

Classifier

ASD probability head (BCE loss) — the direct autism-association signal.

02

Subject encoder

256-dim embedding used for cohort stratification and subtype clustering.

03

Reconstructor

Rebuilds the 4,950-dim connectivity — its error is the QC / atypicality signal.

227.7M
Trainable parameters
1,545
ABIDE subjects
36
Clinical sites
63%
Validation accuracy

Honest disclosure

What Cortex is, and what it is not. Read this before trusting a number.

Not for clinical diagnosis

~63% validation accuracy on held-out ABIDE subjects is useful for research, not diagnosis. ADOS-2 / ADI-R remain the clinical standard.

Cohort is ABIDE I + II

Predominantly male (~91%), mostly North American / European sites, ages 5–64. Generalization to underrepresented populations is not validated.

Sites harmonized, not erased

Residual site bias may remain after harmonization. Use caution when applying to new acquisition protocols.

ASD vs TD is a simplification

Real neurodiversity is spectral. The 256-dim embedding reflects that far better than the binary classifier logit.

Start with your data

Upload connectivity features. Explore subtypes, flag outliers, discover biomarkers. No signup.