Oleg Zabluda's blog
Thursday, July 26, 2018
 
All-optical machine learning using diffractive deep neural networks
All-optical machine learning using diffractive deep neural networks
"""
Here we introduce an all-optical deep learning framework, where the neural network is physically formed by multiple layers of diffractive surfaces that work in collaboration to optically perform an arbitrary function that the network can statistically learn. We term this framework as Diffractive Deep Neural Network (D2NN) and demonstrate its learning capabilities through both simulations and experiments. A D2NN can be physically created by using several transmissive and/or reflective layers, where each point on a given layer represents an
artificial neuron that is connected to other neurons of the following layers through optical diffraction (see Fig. 1A). Following the Huygens’ Principle, our terminology is based on the fact that each point on a given layer acts as a secondary source of a wave, the amplitude and phase of which are determined by the product of the input wave and the complex-valued transmission or reflection coefficient at that point. Therefore, an artificial neuron in a diffractive deep neural network is connected to other neurons of the following layer through a secondary wave that is modulated in amplitude and phase by both the input interference pattern created by the earlier layers and the local transmission/reflection coefficient at that point.
[...]
This D2NN design [can be] physically fabricated using e.g., 3D-printing, lithography, etc. To experimentally validate [...] we used terahertz (THz) part of the electromagnetic spectrum and a standard 3D-printer to fabricate different layers [...] we focused on transmissive D2NN architectures (Fig. 1) with phase-only modulation at each layer. For example, using five 3D-printed transmission layers, containing a total of 0.2 million neurons and ~8.0 billion connections [...] This D2NN design formed a fully-connected network and achieved 91.75% classification accuracy on MNIST dataset
[...]
using different 3D-printed layers, we experimentally demonstrated the function of an imaging lens using five transmissive layers with 0.45 million neurons that are stacked in 3D. This second 3D-printed D2NN was much more compact in depth, with 4 mm axial spacing between successive network layers, and that is why it had much smaller number of connections (<0.1 billion) among neurons compared to the digit classifier D2NN, which had 30 mm axial spacing between its layers.
"""
http://science.sciencemag.org/content/early/2018/07/25/science.aat8084
http://science.sciencemag.org/content/early/2018/07/25/science.aat8084

Labels:


| |

Home

Powered by Blogger