What if a neural network only needed 1.58 bits per parameter — and still stayed accurate?

That's the idea behind BitFace, recent research I led: a ternary Vision Transformer (each weight takes just 3 values: -1, 0, or +1) built for face recognition on low-power edge devices.
Why it matters: most high-performing vision models are too heavy for edge computing — security cameras, embedded devices, IoT. We end up sending data to the cloud, with all the latency, energy cost, and privacy tradeoffs that brings.
By pushing quantization to the extreme, we drastically cut:
→ memory footprint
→ energy consumption
→ inference time
... without sacrificing recognition reliability.
It's one more step toward AI that runs directly on-device, not just in a datacenter.
Preprint is live on SSRN — link in the comments if you're curious. Would love to hear from anyone working on edge AI or model compression!