Neural digital twin

An in-silico foundation model that predicts a brain region's responses to arbitrary inputs — an encoding-model foundation model — trained across many recordings so that it generalizes to new stimuli and, increasingly, to new individuals. Large models of mouse visual cortex are the clearest example: trained on responses to natural movies across many scans, they predict activity to novel stimuli well enough to run experiments in silico, search for optimal (maximally exciting) stimuli, and probe tuning without touching the animal.

The twin is the encode direction, the mirror image of a decoding foundation model. That direction is exactly what a write-in prosthesis needs: to know which stimulation pattern will evoke a target percept, you need a forward model of how the tissue responds. Digital twins are therefore a natural companion to visual and other neuroprostheses, and a testbed for hypotheses that would be too expensive or slow to run in vivo — with the standing caveat that a twin is only as good as the stimulus distribution it was trained on.