DEEP NETWORKS WITH ULTRA-FAST PHOTONIC TRAINING FOR INSTANTANEOUS DIRECT IMAGE CLASSIFICATION
Abstract
Deep learning using deep neural networks is increasingly in demand for many applications from image and speech recognition to artificial intelligence. The state-of-the-art method for implementation of such networks is to utilize advanced parallel graphics processing units (GPU) where despite excellent performance, it suffers from high power consumption and long processingtime for complex multi-layer networks. This project introduces two photonic mm-wave architectures for implementation of deep neural networks for image recognition. In the first system, the images are taken using an array of pixels and directly processed in the optical domain for both learning and classification phases with part of the processing performed in mm-wave blocks. Thisproposed architecture directly processes the raw optical data (without photo-detection/digitization) with orders-of magnitude less processing time than the state-of-the-art, orders-of-magnitude lower power consumption, and scalability to complex practical deep networks. In the second system, the array of the input pixels is replaced with a fast-steering two-dimensional multi-beam opticalphased array enabling applications such as direct 3D image classification, phase contrast image classification, and tracking and classification of fast moving objects. Both proposed photonic deep networks are highly reconfigurable as the number of layers and neurons per layer can be adjusteddynamically. Under this program, both architectures will be designed, implemented, tested, and used to demonstrate training and classification of complex images. - - - - - - -
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 25, 2019
- Source ID
- N000141912248
Entities
People
- Firooz Aflatouni
Organizations
- Office of Naval Research
- United States Navy
- University of Pennsylvania