Dida is an interdisciplinary team of mathematicians, physicists and engineers who develop tailor-made software applications for companies. With the help of machine vision, structures in image and text can be recognized that people do not recognize - with use cases in the insurance, e-commerce and travel, real estate and health sectors. #ki_berlin spoke with dida’s CEO Philipp Jackmuth and CTO Lorenz Richter about pattern recognition, data sharing and public AI discourse.
You are working hard in the area of process optimization and how deep learning can be used for this. Machine vision is one of your hobbyhorses. What does it mean?
Lorenz Richter: Machine vision is all about using algorithms to recognize and analyse patterns in images. Classical tasks are the classification of objects or the segmentation of images into given units. The main feature of the deep learning approaches is that the corresponding algorithms can generalize. This means that they can also handle data that they have not seen before. No stubborn set of rules is implemented, but attempts are made to represent a semantic structure by means of complex nonlinear functions (so-called artificial neural networks). The progress made in this area of machine learning over the last five years has been enormous. Meanwhile, we can start from the rule of thumb that anything a person can recognize in an image can also be recognized by an algorithm. To name a specific application: in a current project for the German Weather Service, we are identifying emerging thunderclouds by analysing satellite images. This would not have been possible a few years ago.
Which topics do your machine vision solutions deal with, and can they be transferred to other areas?
Philipp Jackmuth: The beauty of deep learning is that it is thematically to some extent universal. Whether the recognition of dogs and cats in images or the analysis of complex satellite data - the underlying machine learning models function in a very similar way. So we are not focused on specific industries, but rather on the general mathematical modelling of problems in the areas of pattern recognition in images (machine vision) and pattern recognition in texts (natural language processing, NLP). Since a lot is currently happening in these research fields, our customers benefit greatly from the fact that we are strongly linked to scientific research and invest a lot of time in internal training.
The recognition, processing and use of machine data is one thing. You are talking about natural language processing, which affects practically all areas of the (working) world. How does this type of pattern recognition work?
Lorenz Richter: Natural language processing is about analysing unstructured text, i.e. usually continuous text as it appears in letters or in interviews like this one. In contrast to images, here we are dealing with sequential data in which the order plays a role. In addition, it is much less clear how to encode the data in a mathematical model. With images the pixel values are used as input variables, for texts the type of coding is the first step in the design of a working algorithm. In contrast to image data, so-called recurrent neural networks have proven to be effective here, in which feedback loops can exist within the model in order to take the sequential structure into account. The large presence of unstructured text and the associated potential to profitably incorporate this data into internal processes is certainly exciting.
Dida is very interdisciplinary. Your team consists of mathematicians, physicists and engineers. How do you assess Berlin as an AI location and the link between research and industry in the city?
Lorenz Richter: We indeed believe that, in addition to the current trend of using pre-implemented "black box" models, it is essential to refer to the mathematical foundations of machine learning. In our sometimes highly specialized applications it is important to understand the details - mathematics and physics are two grateful disciplines which are indispensable along with the engineering aspect. With its three large universities and cutting-edge research as well as very good education in the quantitative areas - for example at the Berlin Mathematical School or the new Math+ excellence cluster - Berlin is certainly well positioned. Machine learning is also becoming represented more and more in the university landscape. In addition, there is a lively start-up scene and the establishment of AI stars such as Amazon and Google, which are certainly good for the city.
The USA and China in particular are making immense efforts in the field of AI. What must happen so that Germany does not fall behind in an international comparison?
Philipp Jackmuth: It is certainly important to consistently promote basic research and companies with innovative AI ideas. Accordingly, we welcome the funding currently being offered by the Federal Ministry of Education and Research in the area of artificial intelligence. Furthermore, government agencies in particular should make as much data as possible freely available - keyword "Open Data". This is also desirable from the point of view of democracy and transparency. Here, too, we see commendable approaches, for example within the framework of the Copernicus project of the European Space Agency (ESA). Last but not least, public dialogue on new technologies should also be promoted in order to provide the interested population with good, non-sensational information and to dispel unrealistic fears.
How do you see the willingness to innovate and invest in artificial intelligence in Germany?
Lorenz Richter: Compared to the USA or China, we certainly have some catching up to do here. A certain scepticism towards artificial intelligence is surely appropriate, but at the same time it must be recognised that the potential for progress is enormous and that certain efforts are necessary for progress to be made. We are optimistic that a lot will also happen here in Germany in the coming years.
Social attention for artificial intelligence has grown rapidly and public discourse fluctuates between fear and anticipation. How do you see the opportunities, potentials and risks?
Philipp Jackmuth: I believe that public discourse is very important, not only in the field of artificial intelligence, but also in any kind of technological progress. The fact is that we humans are developing ever stronger tools. These tools are in themselves value-free, and it depends on the moral judgment of the person who uses them. An old example: I can use a hammer to hang up a nice picture, but I can also hit someone on the head with a hammer. It's similar with artificial intelligence. We need certain social rules that tell us how to deal with this technology. These rules should be worked on in particular by the relevant experts, but in a transparent way and with due regard for public opinion. And it is precisely this kind of discourse that the government should be promoting - not least in order to counteract the sometimes fear-promoting, omnipotent PR cries of certain market participants.
Thank you very much for the interview.