My team and I propose separating the encoder from the rest of the model architecture: 1. Deploy a lightweight encoder on the wearable device's APU (AI processing unit).

“Mu is built with only 330 million parameters, but its performance rivals much larger models,” said Microsoft. It uses a transformer-based encoder-decoder architecture, which processes input ...