Neural Mediapredicted cortical response

About

What Neural Media does — and what it doesn't.

Neural Media runs a short YouTube video segment through Meta FAIR's TRIBE v2 model to predict the predicted average cortical response across the 720 subjects TRIBE was trained on. The four sections below spell out exactly what the numbers in this app mean — and, just as importantly, what they don't.

What it DOES measure

  • Predicted BOLD fMRI response on a 20,484-vertex cortical surface, averaged across TRIBE v2's training subjects.
  • Predicted engagement in well-localized cortical regions: visual cortex, auditory cortex, language network, face-selective regions.
  • How predicted engagement rises and falls across the segment — a per-region time-course, not just a single number.
  • Differences between clips — compare what you paste against the curated gallery examples.

What it does NOT measure

  • Your individual brain. The model predicts an average, not a person-specific response.
  • Subcortical reward / addiction circuitry. Nucleus accumbens, VTA, amygdala, and friends are not in the output space.
  • Cumulative effects, habituation, or neuroplasticity over time.
  • Subjective states — aesthetic preference, confusion, cognitive overload, memory encoding, satisfaction.

License — CC-BY-NC 4.0

Neural Media is distributed under Creative Commons Attribution-NonCommercial 4.0, inherited from Meta FAIR's TRIBE v2 weights. Personal, research, and educational use is welcome. Commercial use is not.

See creativecommons.org/licenses/by-nc/4.0 for the full license text.

Privacy

  • No account, no sign-in, no tracking — paste a link, get a prediction.
  • Analysis runs on a cloud GPU that fetches only your chosen segment to a temporary directory it deletes after inference — the full video is never stored.
  • Only the small result (the predicted region activations) persists — in AWS S3 / DynamoDB — and auto-expires.
  • No analytics SDKs, no telemetry. The curated gallery is precomputed static JSON: no inference, no network.

Reproducibility envelope: every inference run logs model id + version, random seed, preprocessing params, configuration hash, and a wall-clock UTC timestamp. The mock backend obeys the same envelope so the contract is exercised end-to-end before TRIBE itself arrives.