Scientists at Samara University named after Korolev have developed optical neural networks that will make drones faster, more autonomous, and more energy-efficient. The technology is being created for onboard artificial intelligence of UAVs. The development increases data processing speed and reduces energy consumption, which is critical for autonomous systems.
Researchers have proposed a new approach to building optical transformers — key blocks of modern AI models. These neural networks analyze images and speech using streams of photons and optical components. Unlike digital solutions, optical circuits work faster and consume less energy, but previously lagged in accuracy.
The team solved the problem through a new optical matrix multiplication scheme. The model was integrated directly into the neural network training process, allowing the system to compensate for computational errors without losing overall accuracy.
Experiments confirmed the calculations on language models, computer vision systems, and applied AI tasks. A miniature diffraction element, used in ultralight optics for nanosatellites, was added to the scheme.
