What is neural engineering

"We simulate the structure of neurons"

The human brain consists of around a hundred billion nerve cells, the neurons, which are linked by so-called synapses. In some respects, this architecture is superior to even the most modern computers. Therefore, scientists try to mimic the behavior of neurons and synapses. Most of the existing approaches to such neuromorphic networks currently rely on electrons. A new system, on the other hand, uses light particles to transmit, store and process data even faster. The researchers are now presenting their light-based computer chip with four artificial neurons and a total of sixty synapses in the journal “Nature”. Welt der Physik spoke about this with Wolfram Pernice from the University of Münster.

Wolfram Pernice

World of physics: why are neural networks interesting?

Wolfram Pernice: Neural networks are particularly strong at recognizing all kinds of patterns. These can be images, sounds, languages ​​or abstract mathematical structures. Our human brain is always doing incredible things in this regard - with extremely low energy consumption. The idea now is to use the high speed of optical data processing for such tasks.

What is different about optical data processing than conventional electronic data processing?

In a normal computer, the data memory and processor are separate. The flow of data between these two components represents an important hurdle and brake on computing speed. Optical neural networks, on the other hand, only work with light and not, as normally, with electrons. This means that the computation process can be run through at the speed of light. Such circuits are not as easy to program as a computer. However, they could be used for special purposes, especially in pattern recognition. We have now implemented a neural network on a chip that works with light.

How did you do that?

Our innovation consisted in using a so-called phase change material, the optical properties of which can be adjusted using laser pulses. Such materials can be found in rewritable DVDs, for example. Depending on how long and how strong a laser pulse works at a certain point, the material switches between a disordered state and a crystalline one - in which the atoms are regularly arranged. As a result, its optical properties alternate between transparent and opaque. With our optical circuits, we simulate the structure of neurons in principle.

How can the behavior of nerve cells in the brain be simulated with an optical network?

Neural network chip

We simulated the synapses with a thin layer of phase change material, which allows more or less light to pass through depending on the strength of the connection. Our neurons consist of a so-called ring resonator, whose behavior we can also adjust with a phase change material. In this way, the behavior of the entire system can be set purely optically in a first step. For the "training" we were able to use established methods that are also common for simulations of neural networks on conventional computers. The fully trained system then also provides a purely visual answer to a question we ask it - for example: What kind of letter is that?

How big would such a network have to be in order to cope with interesting tasks?

At the moment we have sixty “synapses” that serve as an entrance and receive image information. These synapses are linked to four “neurons” that ultimately tell us which of the first four letters of the alphabet is present. With more components, we can map the whole alphabet. For practicable applications such as speech recognition, we would need a large number of components - in the range of a few thousand. This is hardly feasible with the manufacturing facilities at a university. But we already use standard processes that are also common in the chip industry. With the appropriate plans, we can implement much larger circuits in the future. However, it is difficult to predict today whether such chips will one day be found on a cell phone.