Prof. Dr. Alexander Löser, Beuth University of Applied Sciences

16 August 2019

"If our established industry doesn't react now, this could lead to bad dependence on uncontrollable monopoly corporations."

Prof. Dr. Alexander Löser is founder and spokesman of the Data Science Research Center at the Beuth Hochschule für Technik (University of Applied Sciences) in Berlin. Seven professors and many doctoral students are working there on methods for deep learning, text, image, mobility or educational data mining. As a consulting expert, he supports Zalando SE, IBM, eBay/mobile.de, MunichRe AG, Krohne Messtechnik as well as three start-ups, and is an expert for EU and Federal ministries such as the BMWi (Federal Ministry of Economics and Technology) or the BMBF (Federal Ministry for Educa­tion and Research) Lernende Systeme“.

You are the founder of the Data Science Research Center at the Beuth University of Applied Sciences in Berlin. In which areas and in which range of topics is research being carried out?

One important common core topic of our basic research is deep learning for very differ­ent modalities such as image, text, tables or graphs. The neuronal representation then allows us to improve the quality with even more varied training data or to create images between two modalities. For the EU project FashionBrain, among others with Zalando, we did exactly that: we linked persons, companies and products which exist in text - for example in blogs - for representation in a database. All this also took place in the main memory database of MonetDB; this has the advantage for the customer that he does not have to buy and maintain another system. For this, there was a Best Paper Award in Kyoto at the IEEE Big Comp 2019.

However, the linking problem also exists in medicine, i.e. when things are to be depicted in medical letters, X-rays or guidelines for ICD-10 codes (note: International Statistical Classification of Diseases and Related Health Problems). So our basic research almost always has a specific purpose, we believe like basic research for AI in industry.

Current partners are therefore mainly from the medical and diagnostic sectors, such as Ada Health, Charité Berlin or Helios, Siemens from supply chain management and Zalando from retail. We are also active in educational data mining and in the optimisa­tion of urban traffic using methods of machine learning.

One of the projects you are involved in is the MACCS Project. What is it about and what are your goals for the next three years?

MACCS was a BMWi project from 2016 to 2019 with the Nephrology department of Charité, SAP, the start-up SmartPatient and other medical partners. The doctors first want to have the best possible "picture" of the patient before they exclude or make diagno­ses or even suggest therapies. Doctors call this process differential diagnosis. In nephrology, this integrated "image" is made up in particular of data from doctor's letters, but also from vital data, data on medications taken and data from blood or urine samples. Here, too, the aim was to put this "picture" from these multimodal representa­tions into a tangible picture for the doctor. We have developed the neuronal method TASTY, which recognizes symptoms/diseases while a doctor's letter is being written and then assigns ICD-10 codes or their textual representation in Wikipedia. But the key feature is that we could achieve this functionality with the same programming code and only minimal adjustments to parameters with completely different training data in sev­eral languages and completely different domains (medicine, fashion or news). This was already relatively new in 2016. Later, we developed SECTOR, which classifies individ­ual sentences and sections in a document on hundreds of topics. This allows a doctor to make wonderful enquiries, such as: "Give me therapy recommendations for disease xyz" in his archive of tens of thousands of doctor's letters, and he gets only the relevant sentences. This Smart-MD work was presented as a demo at the ACM WWW 2018, and we were able to show the complete analysis at the ACL in Florence in 2019.    

Currently, our vision is to link AI and health even more closely. For example, I am a men­tor at BIH, the Berlin Institute for Health Research, with a focus on translation and precision medicine. Our team is often at Charité; in particular the cooperation with the doctors is a big advantage and the topic fulfils us. We all want to grow old and stay healthy while enjoying life. Smart AI in the health sector, especially in clinical diagnosis, can offer a lot of support. We have now continued this cooperation in a new BMWi pro­ject.

I would like the important players in the field of AI and health in this city to work even more closely with BIH. We should create even more stable structures here. I see the Data Science Research Center at Beuth University as an important partner.

With the development of social media platforms, online debates are more and more characterized by hate comments, slander and threats. With the joint project NOHATE of the FU Berlin you are trying to counteract this trend with the help of data analysis and pattern recognition. How do you tackle this difficult topic?

We deal with the topic like everyone else: as researching, curious engineers! Techni­cally, this is first of all a relatively manageable classification problem: there are a num­ber of letter sequences in the posting, its context in other postings and, for example, the author's context. Thus we maximize the probability of predicting the right class (here: hate speech or not).

But then it didn't get that trivial. Often only little training data with the truth is available, in other words data with which a forum moderator said: we do not want to have these post­ings. So it is not "Big Data", but often only "Small Data". Then the few assessments of the moderators are anything but homogeneous; they already vary a lot within the forums of a publishing house. With forums of two publishers it becomes quite difficult. Each publisher has its own culture and its own understanding of how much hate speech should be deleted and how much should be published, in the sense of freedom of expres­sion. These models must therefore be able to achieve the socially important balance between the basic democratic understanding of freedom of expression and the deletion of insults as easily as possible. We then solved the problem with approaches of transfer learning and ensemble learning. In transfer learning, a very large language model of the target language is used, which learns by itself which word or sentence could come next. The current models BERT, ELMo, GPT-2 or BigBird are trained on hundreds of millions of sentences and have many millions of parameters. These are very powerful models which can also encode many deviations and variations of lan­guage.

We are currently working with publishing partners at heise, golem.de, Gute Frage, Welt and other publishers on the next version of the models.

Through your collaborations with companies and research institutions, you are very well acquainted with the interface between science and industry. How do you rate Berlin as a location and the link of the two areas in the city?

I came back to Berlin in 2009 from Silicon Valley, first via a small detour to SAP Research. This was a bet that the city would develop well in science. Today I have to laugh and am happy that I "won" this bet.

In my opinion, Berlin is one of the most important science cities in Europe, if not the most important. The physical proximity between industry, investors and academic research is particularly important: in summer, our team can reach our industry partners, other research groups and the federal ministries by bike; in winter it costs 2.80 euro with the BVG, the public transport company. That's great. Numerous well-known companies have settled in Berlin a stone's throw away from Beuth University in Berlin Mitte. We have an English language Master in Data Science, which receives more than 300 applica­tions for 25 places. The students of our Master's programme in Data Science are all highly motivated and very well educated and rightly hope to get good jobs in data science laboratories of the world's leading platforms.

In addition to setting up our Data Science Research Center since 2014, I have been able to work in an advisory capacity on establishing the research and data science depart­ments, for example at Zalando or eBay/mobile.de, for many years. Over the years, very good international teams have grown there and in many other well-known companies and start-ups which I have accompanied. In my opinion, this is mainly due to the good branding of Berlin as a science city.

I believe that this diversity of students and cultures in Berlin in particular is the breeding ground for start-ups. We should deepen this; in particular we should back programmes which also support AI start-ups in the difficult second and third years after their founda­tion. It often takes time to develop the data product and convince the first paying custom­ers. In the area of health, the years for medical studies are added.

In all your projects you work with a wealth of data and information which is proc­essed and used by machine learning processes. How important is the aspect of ethics when it comes to AI, privacy and data protection?

BIAS, i.e. one-sidedness and bias in training data and fairness of algorithms, is an impor­tant topic, not only in our team but also in the entire community. For example, the parole system in American prisons is supported by machine learning: a system recom­mends to a judge whether a prisoner should be released prematurely. In 2017, col­leagues from the United States showed that the system very much tends to favour white prisoners due to design flaws, such as in training data acquisition and prediction. We teach our students and PhD students how to identify these errors and, if possible, elimi­nate them by means of design. Particularly in sovereign tasks, AI should only be used with a great deal of sensitivity - and, above all, should be very well thought out. The air­craft industry has exemplified a culture of error management and, by openly dealing with errors and continuously improving aircraft and processes, has ensured that aircraft are the safest means of transport. For data science and data products, I think the develop­ment of this culture is still ahead of us.

On the other hand, regulations make it extremely difficult for us to design good data prod­ucts in medicine. We virtually need the data of many patients in order to provide the system with as many variants of therapy processes and diagnostics as possible for train­ing purposes. This data is available at the health insurance companies in the high­est ICD-10-coded quality, but German legislation makes it very difficult for us to use this data for our models. Other Northern European countries have already thought further and have central data storages. There are still no clever ideas and no courage, and approaches such as donating data are only at the beginning.  

In countries such as China, the use of pattern recognition, machine learning and AI is in some cases taking on forms which further fuel fears of technology in our society. How do you see the potential here in Germany and in Europe?

As a data scientist, I know that this trained and, compared with China or some countries in North America, rather balanced perception "lands" in the training data for our AI. It may then help us in Germany to recognize fake news rather than less trained eyes in countries which are more clouded by Internet platforms and one-sided reporting media. This well-trained generation is still very active and can shape our society. This is why I do not yet see these developments as in China in our country.      

On the other hand, I wish that the population, politicians and industry beyond the plat­form economy would occasionally have more courage to invest and a greater enthusi­asm for technology. There are studies which show that the market capitalization of the most important American or Chinese players in the platform economy almost reaches the GDP of Germany with more than 3300 billion euro. Unfortunately, according to these figures, Europe is roughly on a par with Africa in this enormous growth segment - SAP SE is the only heavyweight far and wide. This really frightens me as regards the next 20 years. Unfortunately, the development of AI and the platform economy is not a quickly passing hype of some computer scientists which must be sat out, but a topic on which we started too late, and for some industries even the opportunity in the winner takes most markets is already hopelessly long gone and Chinese and American monopo­lies have gained the upper hand. If our established industry doesn't react now, this could lead to a bad dependence of our industry, politicians and population on uncontrol­lable monopoly corporations, which we certainly don't want.

Let’s take a look further: machine learning und human computation. What will the world look like in 2050?

Against the background of the super-fast developments in China and in the platform economy, we ought to have successfully occupied our niches by 2050. In my opinion, these are AI developments in health, B2B platform economics for German industry, ecol­ogy, the energy industry and lifelong learning. In these markets we still have, at least in my opinion, a certain chance to increase our existing substance with AI and to positively shape our culture.

Thank you for your time.