Latent Surface
We are building generalizable neural foundation models to recover subject-invariant representations of brain computation.
Neural data is usually modeled one person at a time. That makes it hard to separate the computations shared across brains from the measurement details unique to each patient, electrode array, and recording session. Latent Surface develops models that align intracranial brain recordings into shared anatomical and functional representations, starting with speech.
A cross-subject neural speech decoder is a first step toward a larger goal: understanding the invariant structure of neural computation. If we can recover that latent surface, we can build stronger AI systems, improve brain-computer interfaces, and study intelligence and consciousness with models that generalize beyond one brain at a time.