Involved in Henk Corporaal's project alongside TU/e are a dozen universities and research institutions and some thirty companies, among them ING, Schiphol, Siemens, Intel, Océ, TomTom and Tata Steel. In total, according to Corporaal, almost seven million euros are involved in his program, four million of which are provided by NWO.

Based on actual situations experienced by the affiliated companies, the program aims to make deep learning applicable for, among other things, automatic visual inspections, tissue analysis, the smart maintenance of equipment, and intelligent hearing aids able to handle noisy surroundings. The intention is for the program to provide Dutch industry with an easily applicable service.

Better than people

Deep learning is a promising  method for realizing self-learning systems. These systems comprise many layers of artificial neurons. Each layer learns certain properties, giving it the capacity to recognize structure in the incoming information stream; for example, vertical lines, rectangles, faces or facial expressions. The first layers learn simple structures, while the later layers build on this basic work to master more complex properties. Deep learning has recently surpassed human performance in applications such as image recognition and the playing of complex board games.

The benefit of self-learning systems, believes Corporaal, is that you don't need to program any algorithms. “The system teaches itself, but that does mean that you have to feed it with a huge number of examples.” Using these examples, the self-learning system allocates values to “between millions and billions” of parameters. “All this requires a massive amount of computing power. Which means that efficiency is absolutely key, certainly if you want to run self-learning systems on, say, smartphones and even smaller devices.”

Hardware

The expertise of Corporaal and his colleagues at TU/e lies primarily in the hardware side of things, he explains. “You first have to think about the network architecture, i.e. how you plan to build the various layers, then you also have to convert that into hardware. Working out how to do that efficiently is my specialism.”

The other programs, to each of which TU/e is contributing, should lead to a new 3D printer for large metal components, microscopy without lenses, new bacteria for the chemical industry, injury-free sport and wearable robotics for people with muscular disorders.

The board of the NWO Domain Applied and Engineering Sciences (AES), the former Technology Foundation STW, is making 21 million euros available for the six comprehensive research programs. Involved companies, civil society organizations and knowledge institutions are complementing NWO's investment with 11 million euros. The total budget will keep 74 PhD candidates and 25 postdocs in work for the coming five to six years.