Performing convolution while streaming data.
Output of the experimental implementation of the convolutional layer using our optical engine when performing classification on MNIST dataset at 20 kHz.
|
|
Demonstration of the
This breakthrough discovery has two far-reaching S&T contributions:
(i) it experimentally demonstrates a prototype with 10x lower latency than the top-of-the-line GPU for the same matrix-size. Such performance is achieved by a the remarkable throughput of up to 1 P-OPS/s at a 8-bit resolution, or 10 P-OPS/s at 1-bit for complex data-sets of CIFAR-10, or MNIST, respectively (OPS = operations per second), and (ii) scientifically it de-validates the assumption that phase-information outweighs amplitude in optical processors for machine-intelligence – a claim made in over 30 years ago and unchallenged until now |
Edge detection using Fourier
We demonstrate an optical processor based on a 4F system which uses low power laser light to perform edge detection at high speed and resolutions, consuming a fraction of power compared to the energy used by current CMOS based technology. In this stage, only the input light signal is modulated with a DMD while in the Fourier plane the signal is filtered using a fixed filter (Gen-0.5)
|
Prof. Volker J. Sorger is the PI if an awarded grant over $3,150,000 from ONRProf. Volker J. Sorger is the PI if an awarded grant over $3,150,000 from ONR. It is a collaboration between GW, UCLA, UT and Omega Optics Inc., with Prof. El-Ghazawi being a Co-PI. The project entitled ‘Photonic Convolutional Processor for Network Edge Computing’ develops an accelerator for feature extraction known as convolutional neural network. Unlike other approaches, this project uniquely combines two million optical parallel channels in free-space with 10's GHz-fast signal processing of silicon photonic chip technology. This one-of-a-kind system enables processing of Peta-bytes (even Exa-bytes when scaled-up) of data every second known only from supercomputers yet at the footprint of a desk rather than a football field. Most remarkably however is, that it can do so at the speed-of-light enabling near real-time information processing, which is not possible in electronics due to parasitic effects of electrons in wires. Such a real-time processor is a sought after technology by the military for timely decision making, but also civilian applications could benefit from it such as autonomous and self-driving or flying vehicles.
|
Copyright 2019 - George Washington University