Pluggable Diffractive Neural Networks (P-DNN): A General Paradigm Resorting to the Cascaded Metasurfaces that can be applied to Recognize Various Tasks by Switching Internal Plugins

The deep learning method, a machine learning technique inspired by the human brain, has applications in various domains, such as image processing, image recognition, speech recognition, and language translation. However, it relies heavily on electronic computers, which have computational limits, and due to the von Neumann architecture, which leads to bottlenecks in performance and high energy consumption. The optical neural networks optimize the light and offer solutions to these problems by enabling high-speed, parallel, and energy-efficient consumption.

The authors have introduced P-DNN as an innovative solution to the reconfigurability problems of ONNs. Unlike the traditional methods that require complete retraining when a new task arises, P-DNN can switch recognition tasks by swapping the pluggable values in the network. This feature enhances the flexibility of the network design while effectively reducing the consumption of computing resources and training time. The researchers have used two-layered cascaded metasurfaces to demonstrate the approach by using handwritten digits and fashion as inputs, respectively.

The P-DNN architecture includes a common preprocessing layer and alternative task-specific classification layers. The system is trained based on the optical diffraction theory, with each layer’s optical neuron represented by meta-atoms in the meta-surfaces. The training phase involves optimizing the parameters of the metasurface components using stochastic gradient descent and error backpropagation methods. The article highlights optimization flow on transfer learning, allowing the system to achieve high accuracy for various classification tasks The article presents results for digits and fashion classification tasks using the P-DNN framework. Both simulation and experimental tasks show high accuracies, more than 90% for both tasks.

The pluggable Diffractive neural networks act as a solution to the limitations of traditional deep learning by leveraging optical neural networks. It can cater to a range of specific tasks, not restricted to classification tasks alone. They offer energy-efficient, high computational systems for real tasks like real-life object detection in autonomous driving and intelligent object filtering for microscope imaging.

Check out the Paper. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 29k+ ML SubReddit, 40k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.

If you like our work, please follow us on Twitter