Optical 'deep learning' for computers tested

Posted Jun 14, 2017 by Tim Sandle
A new approach for the computations necessary for machines to fully utilize 'deep learning' and to take the necessary steps on the path towards artificial intelligence has been devised and it is based on optics.
Close up of a silicon chip
Close up of a silicon chip
Ioan Sameli (CC BY-SA 2.0)
The new approach for deep learning stems from the Massachusetts Institute of Technology (MIT). Deep learning refers to the use of artificial neural networks which contain more than one hidden layer. Deep learning is part of a broader family of machine learning methods based on learning data representations. As an example, most image recognition machines are "trained" via deep learning. In many cases such machines can complete task like looking for cancer in blood better than people. A celebrate example was with Google’s AlphaGo which beat the world number one human Go player at his own game.
READ MORE: Google's DeepMind AlphaGo defeats Go champion
The approach taken by MIT researchers is to use light instead of electricity. This involved creating a computer with a light-based neural-network system in the form of a programmable nanophotonic processor. This works by using multiple light beams directed so that their waves interact with each other. This produces what are called interference patterns, designed to convey the result of the intended operation.
The research outcomes, according to Controlled Environments magazine, suggest that this approach will considerably improve the speed and efficiency of certain deep learning computations; tests also showed such devices would be able to carry out calculations performed in typical artificial intelligence algorithms using less than one-thousandth as much energy.
According to one of the researchers, Professor Marin Soljačić: "This chip, once you tune it, can carry out matrix multiplication with, in principle, zero energy, almost instantly." However, he also notes there is more to do: "We’ve demonstrated the crucial building blocks but not yet the full system."
The research has been published in the journal Nature Photonics under the title "Deep learning with coherent nanophotonic circuits."