Calculating with the fiber optics

The researchers themselves are still puzzling over why the result captured by the camera at the exit of the fiber can then be classified so easily and with a single additional neuronal layer. In any case, there are indications that the transformation is not yet optimal. For example, the performance depends on the exact shape of the glass fiber. So it will change if accidentally bent or coiled differently. Although this offers the possibility of further optimization through trial and error, this instability also poses a problem. »If a commercial product is ever to result from this, the system must be stable enough to ensure reliable results months and years after the training to deliver,« says Moser. The fiber optics, which are still loose at the moment, would therefore have to be permanently integrated into a chip. According to Moser, it is already possible today to write a one-metre-long light guide on the surface of a silicon chip – in the form of a spiral just one square centimeter in size.

More “conventional” optical chips are also making progress

“Optical computing is particularly promising when it is integrated on chips and relies on elements that are already well under control,” agrees Cornelia Denz, who heads the working group for nonlinear photonics at the Institute for Applied Physics at the University of Münster and last year co-authored a review article in »Nature« on optical computing and artificial intelligence. Such systems integrated on silicon chips are already very advanced and are partly manufactured using the same methods that are also used in the production of conventional electronic chips – except that instead of transistors, light guides, beam splitters and other optical elements are written into the silicon . In this way, the advantages of optical computing, such as low losses, high clock rates and the possibility of parallel data processing, can be packaged together in a very promising way. “Therefore, it can be assumed that this will soon also be used in commercial products,” says Denz. »You will probably hear more about it in the next five to ten years. The current start-ups also show that this is possible.«

The chips from the two US start-ups Lightmatter and Lightelligence appear to be particularly advanced. They are based on the same scientific paper published in 2017 by researchers at the Massachusetts Institute of Technology. The two lead authors are now competitors in an exciting race for the new technology. From a purely financial point of view, Lightmatter has taken the lead in 2021 with an additional injection of $80 million from GV (formerly Google Ventures) and Hewlett Packard, among others. But competitor Lightelligence can also refer to a total of 100 million dollars. According to their own statements, both start-ups are about to launch their respective products on the market.

Compared to SOLO, Lightelligence and Lightmatter are relatively close to the electronic models in terms of the architecture of their optical chips, except that they use so-called Mach-Zehnder interferometers as elementary switching elements. These optical components first split a light beam into two parts at their input and combine them again at the output. Because the speed at which the light propagates along the two different paths can be deliberately altered, there is a delay in the recombining of the rays. In this way, the positions of wave crests and wave troughs can be shifted in relation to one another, causing them to interfere either constructively or destructively in the subsequent superimposition. So the device is essentially a modulator that changes the intensity of a light beam.

“Basically, any mathematical operation can be implemented with a network of Mach-Zehnder interferometers,” explains Rolf Drechsler, who heads the Cyber-Physical Systems research department at the German Research Center for Artificial Intelligence and the Computer Architecture working group at the University of Bremen. In recent years, the researcher and his team have developed instructions on how to interconnect optical components for a wide variety of purposes. They used tiny interferometers because they work best based on the current state of the art and are therefore the most promising for future applications. “In principle, it would even be possible to produce a fully functional, optical computer,” says Drechsler. However, many things could not yet be technically implemented in the desired robustness and quality. In addition, a technical application must also be compact and inexpensive in order to be able to use it commercially.

The current start-ups still have to provide the final proof that optics can prevail, at least in special AI applications, against the silicon electronics that have been optimized over decades. But the time seems to be ripe to give artificial intelligence a new impetus with a bit of optics.