frontier models
Four Million Optical Neurons Arrive in Fifteen Minutes
A task-agnostic diffractive encoder creates a huge random feature map, leaving only a small digital readout to train.
Summary
A task-agnostic diffractive encoder creates a huge random feature map, leaving only a small digital readout to train.
Researchers reported a random-projection optical neural network whose diffractive encoder can be printed in about 15 minutes and supplies roughly four million optical neurons at 500-nanometer spacing. A fixed physical layer scatters light into a high-dimensional representation; a digital readout with about one thousand trainable weights performs classification. Reported benchmark accuracy reached 97% to 99%. This is a laboratory demonstration, not a general-purpose optical computer, but it shows that one quickly manufactured optical front end can be reused across tasks.
Why it matters
A task-agnostic diffractive encoder creates a huge random feature map, leaving only a small digital readout to train.
Limits and context
- This is a laboratory demonstration, not a general-purpose optical computer, but it shows that one quickly manufactured optical front end can be reused across tasks.
Key claims
A task-agnostic diffractive encoder creates a huge random feature map, leaving only a small digital readout to train.
Qualification: This is a laboratory demonstration, not a general-purpose optical computer, but it shows that one quickly manufactured optical front end can be reused across tasks.
Evidence: source-2026-07-17-001
Sources
- CAS via Newswise: Random-projection optical neural networksCAS via Newswise · secondary reporting
Corrections
No corrections have been recorded for this story.