19 November 2018

"It is brilliant when we use machine learning for the important things“.

Prof. Dr. Klaus-Robert Müller: Teacher of Machines

It is due to machine learning that he can predict chemical findings, control walking aids by means of his brain-computer interfaces by mind power alone, and is looking for better ways to diagnose cancer together with the Berliner Charité. For more than 25 years, physics graduate and computer scientist, Professor Dr. Klaus-Robert Müller has been committed to make use of machine learning for ‘important subjects’ such as medicine, natural sciences or neurosciences. The professor, who received numerous awards as a top researchers, is working as a Professor for Machine Learning at the TU Berlin and acts as spokesman for the ‘Berliner Zentrum für Maschinelles Lernen’ (BZML) [Berlin Centre for Machine Learning]. Apart from accepting the position as guest professor at various international universities, since 2014 Professor Dr. Müller has also held the position of Co-Director of the ‘Berlin Big Data Center’ (BBDC).

Professor Dr. Müller, you do not like to be called a pioneer of Artificial Intelligence…?  

That is correct. I really do not like the term ‘Artificial Intelligence’ at all. So far machines are not yet intelligent and many classic AT-researchers of the 80’s have not kept their big promise and only a handful have actually delivered really interesting results.  

You cannot be blamed of the latter in respect of your work. You have made an impact on numerous innovations concerning machine learning: The Support-Vector-Machine, with which characteristic patterns can be recognised in large quantities of data, is used - amongst others - in search engines, in the recognition of picture, handwriting- and speech recognition as well as in the Life-Science sector. With the assistance of the ‘Brain-Computer-Interface‘ (BCI) paralysed persons can again communicate with the outside world, because machine learning algorithms recognise a person’s intention by the characteristic pattern of the brain waves. What exactly spurs you?

I want to do things that interest me, together with inquisitive and interesting people.  Machine learning is one of the very rare fields, where it is possible to achieve improved application results directly by a thorough basic research. Even today, I spend nearly 50 per cent of my time doing research work, as I believe that I am well equipped to reason some matters better because of my long-standing experience. Also, I want to show that it is really worthwhile to dedicate one’s energy to really important things. Together with the pathologists at the Charité, for example, I am working at a project, where cancer cells can be identified more precisely, be classified and their destructive effect assessed by means of a self-training picture recognition procedure. The achieved progress is remarkable.  

The question, how learning- and decision processes of Artificial Intelligence can be made transparent and be explained, seems to be a central subject in performing your work. Your ‘Layer-wise Relevance Propagation“ (LRP) is an important development to light up the ‘Black Box‘ of machine learning systems. Can you please explain this system in more detail?

LRP, which we have developed and patented mutually with the Fraunhofer Heinrich-Hertz-Institut, is a method to understand neuronal networks better. In a figurative sense, these are made up of various layers of interlinked, self-learning algorithmic elements, similar to human neurons.  For example, in order to teach such learning system to recognise trees, they are ‘fed’ with pictures showing different trees. However, the tree itself is not marked on the pictures but the whole picture is labelled ‘tree’ or ‘not a tree’.  Gradually, the system bundles all feedbacks and evaluates these, until every tree on every picture is recognised. The LRP looks at these individual decision-making processes in layers backward and at the same time calculates which neurons have made what decisions and how relevant these decisions were for the final result. This is optically demonstrated in a so-called ‘Heatmap’. This map shows which pixels have contributed particularly greatly to class the picture as a tree or not a tree. This method, to be able to interpret results of neuronal networks subsequently, is a decisive step forward, especially since the system can not only be used to recognise pictures but can be made use of universally.  

Concerning areas of application: Are there any areas, where this ‘Explainable Artificial Intelligence’ may become more important than in others and why?

Especially in vital cases, meaning in respect of safety-relevant or medical questions, the users want to understand exactly why a machine learning system has made its decisions. Explaining decisions and understanding which neurons have made what decisions and how strongly the final result is influenced by these decisions, give a meaning to the use of data-driven learning algorithms.  

You apply machine learning also to the so far neglected classic Natural- and Material sciences in general and molecular-dynamic situations in particular representing the basis of many models in chemistry or biology. Why were machine learning processes hardly used in these areas?  

Most of the machine learning processes work with standard algorithms that assume that the quantity of the data to be processed is irrelevant. This does, however, not apply to precise quantum-mechanical calculations of a molecule, where every individual data point is decisive and where an individual calculation of larger molecules can take weeks or sometimes even months. Because of the enormous performance of a computer that is required for this, it was so far impossible to carry out ultra-precise molecule-dynamic simulations.  

That was the situation so far - but thanks to your method, new Natural Science findings can be obtained easier by now. How did you achieve this?  

The trick is not to calculate all of the potentially possible conditions of the molecular dynamics with the machine learning process, but only those that do not result from known physical laws or by applying symmetry operations. These special algorithms allow the process to focus on difficult problems of the simulation instead of using the computer power to reconstruct trivial relations between data points. This demonstrates impressively the high potential of combining artificial intelligence and chemistry or even other natural sciences.

Recently, you have been working at a new research project that was funded by the BMBF [Federal Ministry for Education and Research]. It is all about researching the basics of digitalised product development in the car industry. Can you please tell me more about the Artificial Intelligence Aided x (AIAx) project?

Digital prototypes enable already at this point for a new product to be analysed by means of simulations without actually having to build a prototype. However, every simulation supplies huge volumes of data that have to be currently evaluated manually by engineers in order to recognise and improve deficiencies in the design. This data is to be automatically analysed intelligently and possible suggestions made by special machine learning processes. Above all we will deal with the subjects ‘Efficient Deep Learning’ and the ‘Explainability and Robustness’ of the processes to be developed. By using a proband study, we are going to examine different ways of presenting decision-making processes. In what way can the explainability be depicted? What information is useful? The explainability is an important factor in respect of accepting machine learning processes as in the end the designer is responsible.  

In order to move developments of machine learning forward, especially in the economic sector, you, as a scientist, depend on (private) data. You are fighting for the right of privacy in the net. Why is the latter so important to you?

Privacy is a very precious asset that must be protected. In my opinion democracy does not work without privacy.  

In your opinion, what is necessary to protect private data sufficiently and still continue to develop?  

There are technical algorithmic options with which information can be extracted without having to save everything. Therefore it is necessary to find reasonable regulations where privacy is still protected but enabling technical boundary conditions, such that our data cannot be used in a digital ’Wild West‘ manner.  

For more than 15 years you have been particularly engaged as Professor for Machine Learning at the TU to promote scientific offspring. Successfully. By now a record-breaking 31 of your once doctoral students or post-doctoral students have by now established themselves as full-time professors in the U.S., Europe or Asia. Why is the scientific offspring so important to you?

During my lectures to the students I describe how wonderful it can be to use machine learning for really important matters. For example, to find out, whether blood poisoning has been caused by bacteria or viruses. A large number of my graduates, however, still decided to join the Silicon Valley and start a career there. Others establish their own company, especially round Berlin.  They work in social media or digital marketing. I can understand them as knowledge that we have generated, is unfolded there with immediate impact and income.