The diffractive processor designed with deep learning can perform hundreds of transformations simultaneously
The complexity of computational tasks has increased in the digital age. The power consumption of digital computers has increased exponentially as a result. It is therefore necessary to create hardware resources capable of performing large-scale computations in an energy-efficient and fast manner.
In this respect, optical computers that use light instead electricity to perform calculations are promising. The parallelism of optical systems can help them achieve lower latency, and reduce power consumption. Researchers have therefore explored different optical computing designs.
A diffractive network can be designed using optics and deep-learning to perform complex computations such as image reconstruction and classification. The stack is composed of structured diffractive layer with thousands of diffractive neurons/features. These passive layers control the light-matter interaction to modulate input light and produce desired output. The researchers train the diffractive networks by optimizing these layers’ profiles using deep learning tools. This framework can be used as an optical processing module after fabrication. It only needs an illumination source for power.
Source:
https://phys.org/news/2023-01-deep-learning-designed-diffractive-processor-hundreds.html
Leave a Reply