"Our job is to make machine learning techniques usable"
In the early days, artificial intelligence (AI) was essentially the science of computer programming. Since then, it has developed more and more into an investigation of human thought processes and the attempt to recreate human intelligence by means of machines. In parallel, its use in industry has grown, particularly in the automotive sector.
Self-driving cars, for example, require systems that, like humans, are able to react flexibly and safely in a variety of situations, especially critical ones. For ETAS, this means new business opportunities – not only in product development, but also in advanced engineering. For some time now, we set up a machine learning team to carry out general, non-project-specific research in this area to generate potential openings for the company.
The machine learning team at ETAS
Team leader Holger Ulmer, Chief Expert Machine Learning at ETAS, and Machine Learning Engineer and team member Andrej Junginger, explain together, why artificial intelligence is so crucial for the future of ETAS, where machine learning is already being used, and which areas of research have already produced successful results for the team.
Holger, can you please explain the difference between artificial intelligence and machine learning?
Holger Ulmer: In simple terms, you can say that artificial intelligence is the overarching concept for all the various areas of research concerned with generating computer-based intelligence. In other words, machine learning is a subsection of artificial intelligence and is concerned primarily with the development of algorithms that are able to learn based on data. Then there is a special form of machine learning known as deep learning, which is what our team is principally involved in. Most of the major advances in machine learning in recent years are the result of deep learning techniques. Deep learning is based on the development of artificial neural networks that, similar to the human brain, and as a result of their deep architecture, are able to recognize highly complex structures in a given set of data.
Which ETAS products already make use of machine learning?
Andrej Junginger: Generally speaking, model-based calibration in the area of measurement, calibration, and diagnostics (MCD) is based substantially on machine learning. For example, ETAS ASCMO uses machine learning algorithms to precisely model and then analyze and optimize the behavior of complex systems when there is scant measurement data available. ETAS has been using machine learning for this since 2009. In other words, this method has played a firm role in product development for quite some time now. Likewise, the ETAS ASCMO-MOCA model calibrator uses machine learning techniques for the purposes of data-based modeling. The aim of our interdisciplinary team is to strengthen the ETAS products with additional methods and processes and thus contribute to the long-term success of our company.
Holger Ulmer: Our priority is to come up with solutions to relevant problems, even though we don’t work on a project-specific basis. In other words, our job is to make machine learning techniques usable for concrete products and applications that benefit ETAS. Ideally, we provide users – whether internal or external customers – with expert support as best we can. Our work isn’t supposed to be purely academic; rather, it should result in genuine improvements of products, etc.
Why is it such a crucial topic for the future of ETAS?
Holger Ulmer: For a start, automated processes based on machine learning help us become more efficient, as machine learning tends to get used in areas where machines can take over processes that are relatively easy to learn – simple administrative tasks, for example. Second, when integrated into our products, they add significant value for our customers, as they help them become better at what they do. At the same time, our customers themselves are making greater use of artificial intelligence in their products and solutions. What we’re seeing is a paradigm shift in the development process. For us, this means we need to take these new circumstances into account and respond to our customers’ changing needs. For example, testing and validating machine learning-based functions pose major new challenges, especially when it comes to self-driving vehicles. We therefore use our expertise and offer appropriate solutions to help our customers adapt to these changes.
Can you tell us about some of the team’s successes?
Andrej Junginger: Our most recent success was a collaborative project with Bosch: high-resolution domain translation of video data – i.e., the process of producing a photorealistic rendering of data from simulated situations. This is particularly relevant for the development of driver assistance systems or for autonomous driving, because at present the only way to obtain images is from real test drives. These images are also used to label data – i.e., to tell the system what a tree, a traffic sign, or a crosswalk looks like. Until now, this has to be done manually. With the help of photorealistic rendering, it is now possible to automate this labeling process.
Thank you for the interview!