New Brain-like Photonic Chip Able to “Learn” from Light

A cutting-edge neuromorphic chip uses photons instead of electrons.

Posted May 16, 2019

TheDigitalArtist/Pixabay
Source: TheDigitalArtist/Pixabay

Neuromorphic computing is an interdisciplinary endeavor inspired by biological central nervous system that spans across physics, electronic engineering, photonics, computer science, mathematics, robotics, biology and more disciplines. The biological brain has not only inspired software methods in artificial intelligence (AI), namely artificial neural networks (ANN) used in deep learning, but also hardware architecture. Last week researchers announced the creation of a photonic neural network computing chip that is capable of both supervised and unsupervised AI machine learning, and published their findings in Nature on May 8, 2019.

The research team from the University of Münster, University of Oxford, and University of Exeter, created a hardware chip that processes information like a human brain, complete with artificial synapses and neurons, using cutting-edge photonic techniques. The researchers wrote that they used “wavelength division multiplexing techniques to implement a scalable circuit architecture for photonic neural networks.”

The hardware device consists of four artificial neurons and 60 synapses, arranged in manner where there are 15 synapses per neuron.  The 15 resonators of each neuron required different resonant wavelengths to prevent overlap. The researchers tuned the radius of each ring. A multiplexer consisting of ring resonators is used to “combine the light pulses onto one waveguide and guide it to the on-chip neuron soma” to overcome the challenge that “the input spikes of the artificial neuron are sent on different wavelengths,” reported the researchers.

Next the researchers created simulations to test the hardware’s capability for both supervised and unsupervised machine learning. The reason why they used computer simulations was to overcome the limit of 15 inputs of the hardware device for more robust experimentation. The team created a model with 4096 pre-synaptic input neurons with PCM-synapses (phase change memory synapse).

The researchers ran various simulations for image pattern recognition using 64x64 pixel images, noise patterns, and smiley face patterns. Additionally, the team ran simulations of language detection with a hidden neural network layer to test how well it could determine whether input text was in English or German. The language detection artificial neural network has three hidden layer neurons, and two output neurons, in addition to five input neurons for five vowels.

After training machine learning algorithms by transmitting information through light pulses, the team discovered that the artificial neural network was able to perform pattern recognition.

The researchers reported, “Such photonic neurosynaptic networks promise access to the high speed and high bandwidth inherent to optical systems, thus enabling the direct processing of optical telecommunication and visual data.”

The rebirth of AI from its decades of slumber is largely due to advances in deep learning techniques, the increased availability of big data, the rise of decentralized and cloud-based computing, and the overall increase in massively parallel computing with the GPU crossing the chasm from mostly powering video games to being adapted for general computing purposes, among other recent trends.  Advances in computing hardware further breakthroughs in the software. Now researchers have taken the concept a step further by creating computing hardware that acts in a similar manner to the neurons and synapses of a human brain using all-optical neuromorphic chips that can learn on its own.

Copyright © 2019 Cami Rosso All rights reserved.

References

Feldmann, J., Youngblood, N., Wright, C. D., Bhaskaran, H., Pernice, W. H. P.. “All-optical spiking neurosynaptic networks with self-learning capabilities.” Nature. May 8, 2019.