Computers that function like the human brain could soon become a reality thanks to work carried out on novel, light-sensitive optical fibres. The research, published in Advanced Optical Materials, has the potential for faster and smarter optical computers to be built that would be capable of learning and evolving.
The project was conducted within The Photonics Institute, a recently established dual institute between from the University of Southampton’s Optoelectronics Research Centre (ORC) in the UK, and the Nanyang Technological University’s Centre for Disruptive Photonic Technologies in Singapore.
By using fibre made from a glass that is sensitive to light − known as chalcogenide − the researchers were able to reproduce a range of optical equivalents of brain functions.
‘These materials [chalcogenides] allow not just light to propagate, but can manipulate light within the fibre through changing the material properties at will,’ explained Dr Behrad Gholipour from the ORC and a co-author of the paper.
Chalcogenide materials are amorphous glasses that change their properties in response to varying wavelengths, intensities and polarisations of light – the material’s phase can switch from a glass to a crystal and back again, for example.
In the paper the researchers studied an effect called photodarkening, which involves altering how well the material absorbs light depending on its exposure to different wavelengths. This effect means the efficiency of light propagation through the fibre can be controlled.
‘This [photodarkening] becomes the cornerstone for engineering a lot of these active devices, including neuromorphic systems,’ explained Gholipour speaking to Electro Optics. ‘At a synapse within neural systems you can have a strengthening or weakening in the connection between different neurons in the nervous system. We used the photodarkening in chalcogenides to see if we could get similar effects to the mechanisms taking place in a synapse.’
To simulate a nerve firing, the team first had to mimic a neural resting state − the state before a neuron receives a signal − and then demonstrate the changes in electrical activity in the nerve cell as it is stimulated.
The changing properties of the chalcogenide fibres allow it to perform in the same way as the varying electrical activity in the nerve cell, with the light providing the stimulus to change these properties. This enables switching of a light signal, equivalent to a nerve cell firing.
Neuromorphic computing research has advanced in the last decade, but compared to the human brain, today’s computers are more than a million times less efficient. Simulating five seconds of brain activity takes 500 seconds and needs 1.4MW of power, compared to the small number of calories burned by the human brain.
‘Neural networks are based on massive networks of passive fibres to be able to emulate human brain activity. We have a lot of this functionality inherent in our material,’ explained Gholipour. ‘A traditional neural network would use a huge footprint of different systems to be able to get a synaptic type performance. We can do this with a single fibre, which means we can scale this up.’
The number of neurons in the human brain has been estimated at around 86 billion, and Gholipour stated that he would expect a similar number of these active fibres would be needed to replicate the brain’s processing capacity. However, he said that these fibres could potentially give a faster performance than the brain because the photoinduced effects are much faster than some chemical synapses.
‘That’s where the exciting feature lies,’ he commented. ‘If we start to build a full blown human brain out of these inorganic materials there is the potential to be much faster than a biological brain. Having said that, the main selling point of most neuromorphic research is that, if you look at the amount of energy spent on normal digital computers compared to neuromorphic systems, the traditional binary-based computer is very inefficient. If we could just change our communication protocols reliably to a neural communication protocol and have a medium to be able to implement this, automatically these computers would be orders of magnitude faster.’
In this vision of an optical computer the fibres themselves would both carry the information and also process it. The readout, Gholipour noted, would probably be on a screen, but all the processing would be done by light all within the fibre.
The group is now scaling the fibres to form much smaller waveguides – the researchers are able to fit hundreds of thousands of nodes within a chip the size of a couple of millimetres in diameter. The team is also looking at different materials to increase the photodarkening effect, as well as implementing different devices within the fibre itself.
Greg Blackman is the editor for Electro Optics, Imaging & Machine Vision Europe, and Laser Systems Europe.
You can contact him at greg.blackman@europascience.com or on +44 (0) 1223 275 472.
Find us on Twitter at @ElectroOptics, @IMVEurope, and @LaserSystemsMag.
