Exclusive Talk with Prof. Dr.-Ing. habil. Andreas Maier, Head of the Pattern Recognition Lab of the Friedrich-Alexander-Universität Erlangen-Nürnberg

0
336
Prof. Dr.-Ing. habil. Andreas Maier

Prof. Dr.-Ing. habil. Andreas Maier

Head of the Pattern Recognition Lab of the Friedrich-Alexander-Universität Erlangen-Nürnberg

 

 

 

Asif: Tell us about your educational background and your journey in this highly promising sector.?

Prof. Maier: I am working in pattern recognition since 2005. In the beginning, I was working with speech recognition and speech assessment where we built the first online speech assessment tool for pathologic speech. After my Ph.D., moved on to medical speech processing and joined Stanford University as Post-Doc in the Radiological Sciences Lab where we were developing new image reconstruction methods of which many involved optimization-based solutions. In 2011, I joined Siemens Healthcare where I worked in the Innovation Department developing new image processing and reconstruction methods. Already at that time, we were employing machine learning for medical image analysis. Yet, neural networks were still quite unpopular because of long training times, lack of interpretability, and problems finding the best-matching architecture. In 2012, I joined the Pattern Recognition Lab of the University of Erlangen-Nuremberg, Germany which I am heading since 2015.

Asif: How did you get started in Artificial Intelligence and Neural Networks?

Prof. Maier: Well, I was using neural networks already in 2003-2005 in my Study and Diploma Theses. Already at that time, we were aiming at what is known as „Representation Learning“ in speech data. However, due to the long processing time and shallow networks, we found that linear analyses such as principal component and linear discriminant analysis were obtaining a similar performance at much lower complexity. Note that we used fully connected three-layer-networks way back then. Many scientists obtained similar observations and as such neural networks became increasingly unpopular and were replaced by other methods that were more appealing. First Support Vector Machines were very successful, followed by ensembling techniques such as AdaBoost, and finally Random Forests. I got into neural networks again in 2014/15, as we observed these tremendous results found in computer vision and deep reinforcement learning. One method became quite well received in which we developed an artificial agent that would localize organs in a CT dataset mimicking the procedure of a radiologist that follows other anatomic details until he reaches the point of interest. This approach turned out to be faster and more reliable than any of the competing state-of-the-art approaches. Another noteworthy development is our work on known operators that can be embedded into neural networks. We showed that inclusion of any known operator will reduce the maximal training error bounds, the number of unknown parameters to be trained, and therewith the required amount of training data. It enables us to blend neural networks with any other signal processing technique. The most prominent example is that we can solve CT reconstruction within this theoretical framework very efficiently and that we could derive new previously unknown reconstruction formulas. So this theory is very useful if you need to incorporate physical prior knowledge into a neural network.

Asif: Many people believe that Artificial Intelligence could bring the apocalypse, what are your views?

Prof. Maier: Well, I do not really think that any of the things we are doing right now are really „intelligent“. I would rather refer to what we are doing now as „machine learning“ rather than AI. What were are really good at right now are perceptual tasks in which we need to determine powerful features to assign an observation to a certain class. To some extent what we are doing now reminds me of correlation. We are able to identify relations between complex patterns. What we are still lacking here is the understanding of the underlying factors of cause and effect. Yet, this is tremendously useful to solve common everyday tasks and will yield stronger and more helpful machines and algorithms in our life.

Still, there are some things that can be constructed with this that are extremely terrifying such as AI weapons that can aim and fire on their own. I would put those into the same category as spring guns, land mines, and nuclear bombs. Obviously, such technology and development thereof need to be prohibited.

In terms of the construction of an AI that learns on its own and will at some point overthrow its creator, I believe that we are still missing many steps. Hence, I do not consider this as an immediate danger. However, we need to be careful with any new scientific discovery as it can be used for the good or the bad. The same is also true for other fields of science. If we will be able to cure cancer with advanced genetics, we probably can also use the same technology to create a terrible disease.

Asif: What do you see as the most difficult-to-overcome limitations for AI?

Prof. Maier: Right now, I believe that trust and interpretability of neural networks is one of the most pressing issues. There are some interesting approaches out there such as the integration of known operators as well as Bayesian neural networks that intrinsically produce confidence maps for their predictions. For autonomous driving, you probably want to work on the visualization of the intent of the car and where it will attempt to go next. Maybe augmented reality is a good approach to do so. Yet, there are still open issues such as the existence of adversarial examples that mislead classifiers. They have might be used to hack autonomous cars to mislead them into accidents. Hence, also security plays a vital role in gaining this trust.

Asif: What advice would you give to startups when starting their journey in AI?

Prof. Maier: This is an excellent time to drive your idea towards success. If you think about starting a company, the time is now. However, be careful with who you trust in developing machine learning algorithms. I also feel, as there is a lot of money involved, that there are many „jumping onto the AI train“ that do not really have an in-depth understanding of what they are actually doing. If you have somebody like this as a consultant, you may end up with an overfitted unusable system. So my advice would be to develop demonstrators that work on unseen real-life examples. If you can solve this, you have created a convincing system.

Asif: What advice would you give to students when starting their journey in AI?

Prof. Maier: Take a deep learning class that will appear in your diploma/degree. As far as I have seen this is very beneficial for getting hired by large tech companies. Half of the students that we trained in deep learning got hired by really big tech and automotive companies. I have feeling that people in the HR departments are looking for this keyword on the official certificates.

Asif: Please tell us about your online lectures on AI and how will it help the students?

Prof. Maier: We offer free online recordings of our university. They are accessible to everyone on the following website:

We also offer free Jupyter Notebooks to train implementation in Python:
And last, but not least, we just published an introduction to deep learning under Creative Commons 4.0 attribution license:

Feel free to reuse figures, examples, and text!

Asif: Which books/podcasts/papers have influenced your thoughts the most?

Prof. Maier: I don’t think I can boil this down to few resources. Obviously, there are good textbooks like the one by Ian Goodfellow. The important ones, you can find on MarkTechPost’s list. Other than that, I use twitter quite a lot recently and follow several interesting resources. You can check my twitter account @maier_ak to see which resources I am following. There are also a couple of political accounts that I follow. Note that I do not share the views of several of them.

Asif: What are your views about MarkTechPost?

Prof. Maier: MarkTechPost is an excellent resource as well to inform yourself about the newest trends in technology. When I studied your website, I was impressed with the broad range of topics that you are investigating.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

This site uses Akismet to reduce spam. Learn how your comment data is processed.