A technology that works like a brain? In more complex applications, computers today quickly reach their limits. One of the reasons for this is that their computing units and data memories are separate by concept. As a result, all data has to be sent back and forth. In this respect, the brain itself is many steps ahead of the most modern computers because it processes and stores information in the same place: at the synapses, connections of nerve cells, of which there are about 100 trillion in the brain. An international team of researchers has now succeeded in developing hardware that could pave the way for brain-like computers: The nanoscientists produced a chip on which a network of artificial neurons extends that works with light and can imitate the behaviour of nerve cells in the brain.
The researchers were able to show that such an optical neurosynaptic network is capable of “learning” information and using it to calculate and recognize patterns – just like a brain. Since the system works exclusively with light instead of electrons, it can process data many times faster. “This integrated photonic system is an experimental milestone. The approach could later be used in many areas to evaluate patterns in large amounts of data, for example in medical diagnostics,” says Prof. Dr. Wolfram Pernice, head of the study at the University of Münster (WWU).
The principle presented by the scientists works like this: Optical waveguides are placed on the microchips that can transmit light. The researchers equipped these waveguides with phase-change materials. Such materials are already used today in storage media such as rewritable DVDs. They are characterized by the fact that they change their properties depending on the phase state they are in. The materials change between a crystalline state, in which their atoms arrange themselves in a regular way, and an amorphous state, in which their atoms organize themselves in an irregular way. The phase change can be triggered by light – for example by a laser beam heating the material. “Because the material reacts so strongly and drastically changes its properties, it is well suited to imitate synapses and the excitation transfer between two neurons,” says lead author Johannes Feldmann, who carried out a large part of the experiments as part of his doctoral thesis at WWU.
In their study, the researchers succeeded for the first time in combining a high number of nanostructured phase change materials into a neurosynaptic network. They developed a chip with four artificial neurons and a total of 60 synapses. The structure of the chip, built up in different layers, was based on the wavelength division multiplex technique – a process in which light is transmitted on different channels within an optical nanocircuit.
To test the system’s ability to recognize patterns, the researchers “fed” it information in the form of light pulses and applied two different machine learning algorithms. In the case of the two algorithms used – both for monitored and unsupervised learning – the artificial network was ultimately able to recognize a pattern that had been searched for on the basis of predefined light patterns, including four consecutive letters.
“By working with photons instead of electrons, we can optimally exploit the known potential of optical technologies – not only to transfer data as before, but also to store and process it in one place,” stresses co-author Prof. Dr. Harish Bhaskaran of Oxford University.
In principle, such hardware could be used to automatically identify cancer cells, for example. However, further steps are needed before such applications can be developed. For example, researchers need to increase the number of artificial neurons and synapses and increase the depth of neural networks. This can be done, for example, with optical chips manufactured in standard silicon technology. “This step will take place in the EU joint project ‘Fun-COMP’,” says Prof. Dr. C. David Wright from Exeter University, co-author and head of the Fun-COMP project.
More information: https://www.uni-muenster.de/Physik.PI/Pernice/index.html