An adaptive optical neural network connects thousands of artificial neurons

Modern computer models – for example for complex and powerful artificial intelligence applications – push traditional digital computer processes to their limits. New types of computing architectures, which mimic the working principles of biological neural networks, promise faster and more efficient data processing. Now a team of researchers has developed a so-called event-based architecture using photonic processors where data is transmitted and processed using light. In a similar way to the brain, this enables continuous adaptation of the connections within the neural network. These changeable connections are the basis of learning processes. For the purposes of this study, a working team at the Collaborative Research Center 1459 (“Smart Matter”) – led by physicists Prof. Wolfram Pernis and Prof. Martin Salinga and computer expert Prof. Benjamin Riese, all from the University of Münster – joined the group. It is strengthened with researchers from Exeter and Oxford universities in the UK. This study was published in the journaladvancement of Science“.

What is required for a neural network in machine learning are artificial neurons that are activated by external stimulus signals and communicate with other neurons. The connections between these artificial neurons are called synapses – just like the original biological one. For their study, the team of researchers in Münster used a network of approximately 8,400 optical neurons made of phase-change material along with waveguides, and the team showed that the communication between each of these neurons can actually be stronger or weaker. (synaptic) plasticity), and that new connections can be formed or existing connections removed (structural plasticity). Unlike other similar studies, the synapses were not hardware elements, but were coded as a result of the properties of the light pulses—in other words, as a result of the corresponding wavelength and intensity of the light pulse. This made it possible to integrate several thousand neurons on a single chip and connect them optically.

Compared to traditional electronic processors, light-based processors offer much higher bandwidth, which makes it possible to perform complex computational tasks with lower power consumption. This new approach consists of basic research. “Our goal is to develop an optical computing architecture that makes it possible in the long run to compute AI applications in a fast and low-power way,” says one of the lead authors, Frank Brückerhoff-Plückelmann.

Methodology: Non-volatile phase change materials can be switched between an amorphous structure and a crystalline structure with a highly ordered atomic lattice. This feature allows permanent storage of data even without a power source. The researchers tested the performance of the neural network using an evolutionary algorithm to train it to distinguish between German and English texts. The diagnostic parameter they used was the number of vowels in the text.