| Abstract [eng] |
Deep learning has recently made various breakthroughs in solving complicated inference tasks. Despite digital neural networks becoming more and more popular, they still have some intrinsic detrimental properties. The need for vast energy and computational resources of a digital neural network during inference can be eliminated using optical diffractive neural network (D2NN) design, since only incident photon energy is being used and all computations are done in parallel at the speed of light. While the use of THz waves enables D2NN to be utilized in security, medicine and many other fields. The main task of this work was to design and test the capabilities of physically printed D2NN to classify objects. It was found that diffractive neural network design with 5 hidden layers, distanced no less than 1 centimeter apart showcased the best results. Such D2NN reached an accuracy of 98% when classifying between three chosen objects (a pistol, a knife and a smartphone). While physically printed D2NN performed slightly worse and reached an accuracy of 87.5%. Main factor influencing the drop of accuracy was the hardship in creating high quality smartphone objects, leading to many of them being misclassified. |
