Past the state of the art

We’re excited to share the performance results from our latest set of internal evaluations. Teal’s foundation model has surpassed the 95% accuracy mark on decoding seen images from entirely non-invasive brain data, capable of coverage across a broad swath of the natural language space: animals, plants, food, clothing, body parts, vehicles, technology, household items, and more.

A few key details from these experiments are especially exciting for us.

• We’ve shrunk our inference montage size to just 8 channels, down from standard 32-channel whole-head layouts. This result represents a significant step forward for form factor considerations that have long plagued the possibility of a consumer BCI transition.
• Our model was trained on a corpus containing under 100 hours of data, which—in tandem with accuracy numbers breaking 95%—crucially indicate that Teal’s architecture performs in a highly data-efficient manner. With the current model sitting a healthy distance away from asymptotic behavior on the scaling curve, we are newly optimistic about circumventing long data collection periods and accelerating towards a market-ready product.
• Our data corpus was strictly separated into distinct training and test sets, meaning all reported results were achieved on unseen data. Furthermore, the decoder retained strong performance on generalization tasks designed to pressure test the model’s understanding across different layers of semantic abstraction.

Going forward, we care particularly about the following research directions:

• Cross-subject generalization / reducing new user calibration time
• Minimizing inference latency

We’re sharing these results because we 1) believe that consumer BCIs are close, and therefore 2) want to be transparent about progress.

We love hearing from interesting people. Feel free to message us or check out our open roles.