Review 4th Fraunhofer Alumni Summit

AI already permeates many spheres of life today. At the 4th Fraunhofer Alumni Summit in Berlin on 20 November, many of the attending former employees of Fraunhofer were able to form a comprehensive picture – on tours of the Fraunhofer IPK, the Heinrich Hertz Institute or through the high-caliber speakers from the fields of politics, science, research and economics at the summit's specialist symposium. And of course there were plenty of networking opportunities for the former personnel of the Fraunhofer Gesellschaft.

More than 130 participants came to Berlin on November 20 for the 4th Fraunhofer Alumni Summit at the Fraunhofer IPK and Fraunhofer HHI. For many of the visitors it was a reunion with their former workplace. The Fraunhofer Institute for Production Systems and Design Technology IPK impressed with presentations in the large works hall and with AI-based solutions, for AI-supported gripping in robotic arms for instance. The Fraunhofer Institute for Telecommunications, Heinrich-Hertz-Institut HHI, presented the latest research in optimizing 5G mobile communication technologies with the support of AI. In the HHI's TiME Lab (Tomorrow’s Immersive Media Experience Lab), visitors could also get an idea of how we will experience cinema films in future.

The highlight of the evening was the Summit Night with its presentations and the podium discussion. "In 2019 we are celebrating 70 years of Fraunhofer, our alumni were part of this success and contributed to the fact that in 2019, the Fraunhofer Gesellschaft has grown into an organization with just about 28,000 employees and turnover of around 2.8 billion euros," said Prof. Alexander Kurz, Chairman of the Board and Executive Vice President Human Resources, Legal Affairs and IP Management of the Fraunhofer Gesellschaft. "But figures aren't everything, and today it is mainly about contents and people." In terms of content, Dr. Michael Mertin, Management Consultant, Executive Vice President of Fraunhofer-Alumni e.V. and Fraunhofer ILT alumnus, moderated the evening.

Augmented intelligence und domain knowledge

Dr. Roland Busch, Deputy Chairman of the Managing Board, CTO and Member of the Managing Board of Siemens AG, used impressive examples to illustrate how Siemens AG benefits from the use of artificial intelligence: "Previously we told robots what they should do. Tomorrow the machines will decide for themselves how to do it."

Siemens is thus training industrial robots that independently grip completely different work pieces. Another example is an autonomous tram which avoids accidents with the aid of artificial intelligence. The driver can therefore concentrate on other tasks.

"Everything is becoming digital and we at Siemens believe that the combination of AI and domain knowledge is decisive. Therefore we are also not investing in general AI for the solution of arbitrary problems, but are concentrating on very specific issues," said Dr Busch. An example of this is a pilot project in purchasing at Siemens. The company has around 200,000 suppliers and every year purchases goods and services for around 40 billion euros via around three million orders. "For certain purchasing patterns we use AI algorithms and apply a limit to essential data such as supplier, time of delivery and historical values – in view of the magnitude of the data, the computer can make such decisions better than purchasing personnel. They can then concentrate on strategic topics." Thanks to this augmented intelligence solution, the company could save a significant amount, says Busch.

The fact that a companion approach can lead to better results is also shown in the evaluation of radiology data. "The best results are achieved when AI and the radiologist work together. The system makes a pre-selection and the medics ultimately make the decision," explains Busch. "We don't want to eliminate people, but try to get them to focus on what is essential."

In the case of so-called "Generative Design" artificial intelligence can provide additional value. The company has thus had a high-temperature burner with certain requirements constructed by AI. Inspired by the structure of a fennel bulb, the construction requires 50 percent fewer parts and lower costs by 60 percent. "In this example, the computer has developed a design which an engineer may perhaps never have come up with," said Busch.

The market for artificial intelligence will be worth several billion dollars in the coming years. Nevertheless, Busch also sees constraints, such as data problems for example: highly-paid AI specialists spend 70 percent of their working time on processing data. In many companies – above all in the production sector – there are too few skills in the field of AI. Points such as data security and lack of trust in AI solution conceal additional challenges. Siemens is therefore looking intensively into how AI could be created responsibly, in an explainable and also robust manner.

"Industrial design also requires in-depth domain knowledge," emphasizes Busch. "On the Moscow to St. Petersburg train route, the operators have traveled 13 million kilometers with our trains without a single technical delay. Although there was about as much wear and tear and malfunctions as could be technically expected, our analyses enabled us to predict each imminent failure with a precision of seven to ten days and were able to prevent it though prompt maintenance." However, for this data evaluation very specific knowledge is required. "Therefore it is vital that AI experts work together with the specialist personnel with domain knowledge – a developer cannot work on a solution for a turbine in the morning and optimize a train in the evening."

Trust

It is not only at Siemens that it been recognized how decisive trust in AI solutions is. Prof. Thomas Wiegand, Institute Director of the Fraunhofer HHI, is also intensively involved with this topic: "We're backing big data and machine learning. In doing so it is important that we can trust the result of the AI methods. The Deep Neural Nets (DNN) trained by us use up to 125 million parameters in 32 layers and are therefore very complex. In order to understand these, we have, dedicated thereto, laid the mathematical foundations and provided the algorithms."

The Heinrich-Hertz-Institut is working on a process that explains DNNs. "In this way we can show on the input signal and within the DNNs what primarily led to the classification. We create a heatmap which shows a weighting of the most important to the unimportant features of the input signal. With this it can then be checked, for example, why an algorithm detects a building or a boat." With this method, researchers from the HHI have also examined projects from AI competitions. One winning team was particularly successful in identifying pictures showing horses. As shown in the analysis by the HHI, a particular copyright tag was used for this in most cases, the AI thus being trained to this tag.

This back-projection can, however, also be used in applications in the field of health, such as in analyses for EEG and breast cancer. Incorrect diagnoses in the field of medicine can have fatal consequences, for which reason trust in the reliability of artificial intelligence in healthcare is a major problem which the ITU/WHO focus group Artificial Intelligence for Health (FG-AI4H) is currently grappling with. The aims of this group directed by Prof. Wiegand are the certification and bench-marking of processes and approval proceedings for AI applications in the health sector.

Surveillance capitalism

However, according to Prof. Wiegand, trust in artificial intelligence has a further dimension, as he explained following his presentation: "We are experiencing surveillance capitalism in which billions are being earned on the basis of data, which we have never released for such purposes. In the case of life-style data everyone seems really relaxed. But that is a fallacy. For example, as part of a meeting of the FG-AI4H at Columbia University, we received a (not followed-up) proposal for a project to predict the probability of opioid-dependency on the basis of life-style data."

Personal data stored by social media and other platforms are treated with a certain amount of carelessness. With health-related data on the other hand the utmost sensitivity applies. For Prof. Wiegand this example could be a way to comprehensive data protection: "Perhaps we could use the topic of regulation of health-related data to cure us of the other 'data disease', to regulate this data use and democratically get a grip on it."

Like Dr. Busch, Prof. Wiegand also considers domain knowledge as indispensable. "Recording the data is one thing, the other is annotation by doctors. A topic of the FG-AI4H is an international platform for the collaborative recording and annotation of health-related data."

Data protection as cash cow

On the other hand Prof. Michael Waidner, institute director of Fraunhofer SIT sees no contradiction between economic success and comprehensive data protection, on the contrary: "AI is new form of data processing, but it is still data processing. AI is therefore subject to the same rules. Goals can be achieved even with strong data protection." Protected data have so far not prevented innovation, but it is a matter of being creative with these specifications. "Why not therefore develop a solution like Alexa with perfect data protection," said Waidner provocatively. According to the security expert, higher sales could also be achieved in this way.

But this expert knowledge is also absolutely vital in less specialized AI applications, as is made clear by Jeremie Lecomte, Senior Technical Program Manager for Alexa, Amazon. In reply to Dr. Merten's question as to whether in future only companies that hoard masses of data will be successful, the IIS alumnus declared: "Data without human annotation are practically worthless. At Amazon a very large team processes the data for training the AI solutions. As Dr. Busch said, in the case of problems you have to act very specifically and within a narrowly defined area, the same applies for data. General machine learning or upscaled AI are therefore practically impossible."

Don't rebuild humans!

But what precisely is artificial intelligence? The renowned AI expert Prof. Sami Haddadin of the TU Munich and director of Munich School of Robotics and Machine Intelligence adds that to date there is no absolute definition of [human] intelligence. "We should therefore apply an indicted, anecdotal definition. The more complex a system is and the less it can be foreseen what it does, the more we are willing to ascribe a certain intelligence to a system. Behavior patterns should be subject to blurring and the variation in the surrounds should also be included. Finally, we can also talk about AI if in addition to the learning ability, different environments and variations can be meaningfully combined according to certain metrics."

With Prof. Ina Schieferdecker, head of department at the German Federal Ministry of Education and Research (BMBF), a Fraunhofer alumna was also represented in the podium discussion. "With AI solutions we again have a completely new approach to decision support, decision and intelligence; we don't necessarily have to and should not exclusively try to recreate us humans" said the former institute head of the Fraunhofer FOKUS. "After all, the different examples also show that excellent results can be achieved through interaction between humans and AI solutions – here is also where the greatest opportunities lie in the future."