ZIPBrain: Can EEG Foundation Models Be Faster, Locally Deployable, but Accurate?

Name: ZIPBrain: Can EEG Foundation Models Be Faster, Locally Deployable, but Accurate?
Uploaded: 2026-08-10T09:24:09.225026+00:00
Duration: 1213 s
Description: The hosts discuss ZIPBrain, a training-free token compression module for EEG foundation models. It reduces tokens between attention and feed-forward layers, cutting inference time by up to 41.8% while preserving 99.65% accuracy on TUAB and 97.14% on TUEV at 80% compression. They conclude it enables local, real-time clinical deployment.

The hosts discuss ZIPBrain, a training-free token compression module for EEG foundation models. It reduces tokens between attention and feed-forward layers, cutting inference time by up to 41.8% while preserving 99.65% accuracy on TUAB and 97.14% on TUEV at 80% compression. They conclude it enables local, real-time clinical deployment.

Open the full episode