Researchers have developed a novel method called "optics as code" that utilizes coding agents, powered by large language models, to train optical neural networks directly from hardware measurements. This approach bypasses the need for per-device calibration or complex gradient calculations, instead allowing the agent to learn the physical properties of the optical system and optimize its parameters. The method has demonstrated superior performance in classification and image encoding tasks compared to traditional gradient-free techniques, offering a more cost-effective and scalable pathway for developing optical neural networks. AI
IMPACT This research could accelerate the development of energy-efficient optical computing by enabling more accessible and scalable training of optical neural networks.
RANK_REASON The item describes a new method presented in an academic paper for training optical neural networks. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- backpropagation
- Camera
- coding agents
- Hugging Face
- large language models
- Optical Neural Networks
- optics as code
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →