07. Juni 2018, 09:57 Uhr | Joachim Kroll
Artificial intelligence will drastically change work and products in the future. The speakers at the Artificial Intelligence Forum of WEKA Fachmedien agreed on this. But there is still a long way to go before that happens.
Events on the subject of artificial intelligence are currently held almost every week, expectations are high, promises enormous. In fact, there are already impressive examples of applications with artificial intelligence, for example in medical image recognition for tumor analysis. On the other hand, the automotive industry shows – despite all the performance capabilities of current driver assistance systems – that it will be many years before an »intelligent« vehicle will be able to move through the confusing environment of inner-city traffic independently. The devil is, as so often, in the detail.
This is exactly what the Artificial Intelligence Forum of WEKA Fachmedien was all about. Instead of commercial topics, the focus was on industrial and embedded applications. The three media Computer&Automation, Elektronik, and Elektronik automotive joined forces to create an event with three parallel tracks. In line with the thematic focus of the three media, the tracks dealt with AI in factory automation, embedded applications, and the automotive sector.
The event had previously been introduced by three keynotes. Ralph Bucksch, IBM, presented the state of the art and additionally clarified the meaning of terms that are often confused:
In the second keynote, Volkmar Sterzing explained how Siemens approaches the topic of artificial intelligence in business. The main fields of application are in service (keyword: predictive maintenance), but also the topics of operation in the life cycle cycle, engineering and commissioning are increasingly becoming the AI's field of application.
The third keynote by Dr. Bernd Kosch of Fujitsu was very practice-oriented. He reported on how a manufacturer of turbine blades for wind turbines uses neural networks for quality assurance. Since the turbine blades are only produced in small numbers, there were also only a few data sets for training the neural networks. The method of »Transfer Learning« using a convolutional network nevertheless leads to a high error detection rate and significant time and cost savings.
In the in-depth tracks, such as the embedded track, the technical details were discussed more strongly. A very important point is the training data, with which neural networks are trained and data analyses are carried out. Klaus-Dieter Walter of SSV showed how the data must be preprocessed so that it can even be used for machine analysis.
There are a number of topologies for neural networks: convolutional, recurrent, feedforward, etc. None of the speakers made deterministic statements about which network is suitable for which application or how many layers a network should have. It seems that the search for the optimal neural network is still in the fog and has more to do with trial-and-error than a systematic approach. On the other hand, those who have already gained experience are very reluctant to share their knowledge in order not to disclose this competitive advantage to the public.
Dr. Bernd Kosch, Fujitsu, brought the audience back to the ground towards the end of the event. He grounded the many promises and hype formulas of the AI with a disillusioning statement: »What we call AI today is only the statistical determination of parameters for calculation models«. There is still a long way to go before the systems really acquire context-related skills, "know" the real world, develop an understanding of language and translation and thus make a smooth transition to »natural« intelligence – but the developers are taking great strides in this direction.
The next Artificial Intelligence Forum will take place on May 19, 2019.