3 SHADES OF AI –

5G AND QUANTUM COMPUTING SETTING THE STAGE FOR NEXT GENERATION AI

Inaugural Lecture TU Berlin Berlin, January 22, 2020 Prof. Dr. Sabina Jeschke 1: Intro Breakthroughs in AI The 4th industrial revolution 3 shades of AI

2: AI driven by algorithms Supervised learning Unsupervised learing Reinforcement learning

3: AI further driven by connectivity Basics of 5G 5G changing the mental model of geometrical distances 5G allowing for new AI-driven control models

4: AI breaking grounds via new trends in HPC Moore is hitting the ceiling Alternatives in HPC Quantum Computing as “the new hope”

5: Outro BUA, Berlin and responsibilities Table of Contents of Table CHAPTER 1 INTRO

3 About 2010 – a dramatic breakthrough in Leading to the 4th industrial (r)evolution...

Watson Systems of “human-like” complexity, 2011 leaving the lab environments

Google dimension: complexity level complexity dimension: Car 2012 st 1 About 2010 – a new character of intelligence emerges Everybody and everything is networked, in “real-time”

Car2Infra- structure

Distributed intelligence – new type of intelligence by “real-time-connected brains”

Smart Grid dimension: network level network dimension: nd 2

Swarm Smart Robotics Factory Team Robotics Today – recent progresses in artificial intelligence From “being trained” to “active learning”: getting creative and dealing with uncertainty

Google New trends of neural networks: DeepDream 2015 level With GQNs (generative query networks), AIs autonomously learn to understand the world around – learning from observation like

creativity a child, by imagining and verifying. Scene Edmond de understanding With GANs (generative adversarial Belamy (2018) networks), two neural networks 2018 (discriminator and generator) contest with dimension: dimension: each other. Given a training set, this technique rd

3 learns to generate new data with the same statistics as the training set (e.g., a picture Iamus which fits into a certain group). 2012 The 4th industrial revolution – the “Information Revolution” … is the AI revolution!

New steering and controlling paradigms

Towards intelligent and (partly-) autonomous systems AND systems of systems

Around 1750 Around 1900 Around 1970 Since 2010

1st industrial Power Digital Information revolution revolution revolution revolution Mechanical production Centralized electric power Digital computing and Everybody and everything is systematically using the infrastructure; mass production communication technology, networked – networked power of water and steam by division of labor enhancing systems’ intelligence information as a “huge brain”

“local” “local” to “global” to “global” 3 shades of AI The magic triangle of Algorithms, 5G and Quantum Computing

TODAY TOMORROW DAY AFTER TOMORROW

HPC-driven phase, Connectivity-driven e.g. Quantum Computing phase, Algorithm-driven e.g. 5G phase, e.g. deep learning 3 shades of AI The magic triangle of Algorithms, 5G and Quantum Computing

TODAY

. What are the main algorithmic approaches of today’s AI? . How do they differ, and which one to apply to which kind of problem? . What are applications of today’s AI?

Algorithm-driven phase, e.g. deep learning CHAPTER 2 AI DRIVEN BY ALGORITHMS

10 Supervised – Unsupervised – Reinforcement The three main approaches of machine learning

recognizes patterns recognizes relationships in a data set and learns relationships between inputs and creates between input and outputs in a training data classifications output via feedback Approaches set and makes loops to decide predictions Supervised Unsupervised Reinforcement between about further alternatives and outputs for to steer action new input data data driven data driven trial and error driven (classification/ (clustering/ (learn to react to regression) dimensionality reduction) environment) prognosis analysis acting with decisions Genetic algorithms Decision trees

Monte Carlo Tools k-nearest DBSCAN tree search neighbor Support vector K-means Q-learning machine autoencoder

Deep learning (deep neural networks) Data driven learning – supervised The idea: Learning from existing knowledge

recognizes relationships between inputs and outputs in a training data set and makes predictions Supervised about further outputs for new input data data driven (classification/ regression) prognosis Data driven learning – supervised (@RWTH) … in cooperation with Supervised learning in high-pressure die casting

Can we predict the result of a HPDC (high-pressure die casting) process – by using historical data? – YES, WE CAN! (2015)

Setting: The die casting equipment Add., a “Schrödinger’s cat” in the research wing was seperated phenomenon from the quality check. Thus, the forms were checked with a delay in time and a considerable spacial shift. Under these conditions, IO IO NIO (Outbreak) NIO (Cold shot) „reproducibility“ of the results could NIO (Blowhole) not be reached… Add. steps: integration of weather data and acoustic measurements HPDC process Historic data Prediction model Visualization of measurements process. and Modelling and training prediction quality of data (result NIO|IO with reason) Data driven learning – supervised (@DB) Supervised learning in daily vehicle maintenance, completeness checks

Provisioning Setting: 90% of the processes for the operational maintenance of trains are so- called visual inspections of hundreds of Main- parts. Bottlenecks are time, manpower and Operations tenance plant capacity. Heavy maintenance plants are expensive – camera bridges and neural networks are not. Can we shorten and improve the process of daily maintenance – by Later steps: anomaly detection, see automatically analyzing camera data? – YES, WE CAN! (2019) “unsupervised learning” Data driven learning – unsupervised The idea: Learning by finding similarities

recognizes patterns in a data set and creates classifications

Unsupervised

data driven (clustering/ dimensionality reduction) analysis Data driven learning – unsupervised (@RWTH) Unsupervised learning for laser cutting processes

Can we use unsupervised learning for identifying a group of desired process results in a highly complex process ? – YES, WE CAN! (2016)

Setting: Quality control for automated laser cutting process along parameters coming from 5 optics design parameters (beam quality, astigmatism, focal position, beam radius in x and y directions) and 8 process criteria (roughness values at different depths) Algorithm: k-means clustering groups n multidimensional visualization observations into k clusters, Identification

K=5 clusters...... transferred into parallel coordinates of desired results (i.e. blue cluster) (reduced to the 8-dim. roughness space) Data driven learning – unsupervised (@DB) Unsupervised learning in daily vehicle maintenance, anomaly detection

Setting: Aside from missing parts, the damage of parts is an important issue in operational maintenance of trains. Here, supervised does not do “the Provisioning trick” as the damage patterns and their causes are not known in advance. (Here, 22,000 screws are divided into Main- Operations 50 clusters according to grayscale tenance Example sets distribution.) Algorithm: “k-means clustering” groups n observations into k clusters, Can we shorten and improve the process of daily maintenance – by identification of desired results (i.e. automatically analyzing camera data? – YES, WE CAN! (2019) different causes of damage) Data driven learning – reinforcement The idea: Learning by “trying around”

learns relationships between input and output via feedback loops to decide Reinforcement between alternatives and to steer action trial and error driven (learn to react to environment) acting, with decisions Trial and error driven learning (@world) The maze and more

Can we use reinforcement learning to learn a complex path in a maze without any preknowledge? – YES, WE CAN! (2016)

Super Mario by SethBling, 2015

Setting: An AI is trying to solve a certain super- Mario-Level without any knowledge about the setting. It uses pure reinforcement learning. R- … for learning and executing learning is inspired by behaviorist psychology – complete assembly tasks maximizing the expected return by applying a … for learning of motions sequence of actions at a current state (e.g., Minsky 1954). Algorithm: Neuroevolution of augmenting … as “pro-training” for topologies (NEAT, Stanley 2002), combining human-machine interaction genetic algorithms on top of neural networks. Trial and error – reinforcement learning (@RWTH) http://www.carologistics.org/ Mobile transportation robots for flexible routing

Competitions RoboCup/ Can we use reinforcement learning to coordinate a team of robots to logistics league: cooperatively produce a product? – YES, WE CAN! (2017) 2014: Winner of the World Cup 2015: Winner of the World Cup 2016: Winner of the World Cup Setting: The task is defined only 2017: Winner of the World Cup minutes before the next round. Simulation environments using reinforcement come into play to experiment and to “experience“ decisions and processes. Add.: Totally decentralized, strong cooperation, no ”hard coded components“, intense information sharing, cooperative decision making, re-planning during tasks Trial and error – reinforcement learning (@DB) Reinforcement learning to help coordinate the disposition of vehicles

Day simulation without optimization

= Baseline = smoothed RL / RL = 1st come 1st serve

Provisioning Setting: If the original schedule is not holding, the order of the following vehicles has to be adapted dynamically in a very short time. For a human, the different alternatives Optimized traffic flow can only be rated to a limited extent. Main- Operations tenance Algorithm: trial and error in multi-agent modeling, simulation environment

Can we optimize the throughput of the system in case of perturbances? – YES, WE CAN! (2019) Supervised – Unsupervised – Reinforcement In the mixer!

recognizes patterns recognizes relationships in a data set and learns relationships between inputs and creates between input and outputs in a training data classifications output via feedback Approaches set and makes loops to decide predictions Supervised Unsupervised Reinforcement between about further alternatives and outputs for to steer action new input data data driven data driven trial and error driven (classification/ (clustering/ (learn to react to regression) dimensionality reduction) environment) prognosis analysis acting, with decisions “In the mix” (@world) The creative artificial mind AlphaGo

Can we use “mixed methods” to speed up teach-in processes of robots? – YES, WE CAN! (2016 + 2017)

Setting: Go is originated in China more than 2,500 years ago. As simple as the rules are, the complexity of GO is more than a googol times (10100) larger than chess. “Brute force” does not really work. In March 2016, AlphaGo beats Lee Se- dol (World Champion). Algorithm: Starting with supervised learning, using old games as training data, AlphaGo developed a neural network. Clones of this network play against each other, improving in a “genetic way”, using Monte Carlo tree search. AlphaGoZero skips the supervised part, pure trial and error. It’s even better… “In the mix” (@RWTH) Deep learning/supervised with reinforcement to replace teach-in programming

Can we use “mixed methods” to speed up teach-in processes of robots? – YES, WE CAN! (2017)

Communication via web service connection Simulation environment Q-Learning Environment model (special case of R-Learning) Setting: For lot size 1 teach-in does not work. Robots have to find their way on their own. 𝟏𝟏 𝟏𝟏 𝒙𝒙 𝒙𝒙 𝟐𝟐 Next steps, e.g.: 𝟐𝟐 𝒙𝒙 𝟏𝟏 𝒙𝒙 𝟑𝟑 𝒙𝒙 𝟑𝟑 𝟏𝟏 𝒙𝒙. Improving topology towards Deep neural 𝟐𝟐 𝒙𝒙. 𝒙𝒙 𝟒𝟒 𝒙𝒙𝟑𝟑 𝟒𝟒 𝟐𝟐 𝒙𝒙. network 𝒙𝒙 𝒙𝒙. 𝒙𝒙 .𝟒𝟒 𝟑𝟑 . CNN (convolutional neural 𝒙𝒙. . 𝒙𝒙. 𝟒𝟒 Environmental . 𝒙𝒙. Feature network) . 𝒏𝒏 state 𝒏𝒏 𝒙𝒙vectors 𝒙𝒙 𝒏𝒏 𝒙𝒙 𝒏𝒏 𝒙𝒙 “In the mix” (@RWTH) … in cooperation with Supervised with reinforcement to replace teach-in programming

Can we let a robot perform movements and at the same time adapt to the variable conditions of the real world? – YES, WE CAN! (2017)

An industrial training for „crooked robot... cylinders in the box“ Model predictive regulators Pretraining Setting: An industrial robot with a very elastic joint is to grab objects from a Max. of the process distinct location A and pitch them into a hole at location B. Here, the shapes reward of the objects possess a certain variation – like “real bananas“. - Traditional Optimization of the Dynamics dynamics model model programming is inflexible – slight variations lead to failure of the task. Algorithm: Neural network of 7 (4) layers (3 convolutional layers for vision, Main training process Neuronal Generate training skipped here), 1 expected position layer that converts pixel-wise features to Network data of the model feature points, 3 fully connected layers to produce the torques. “In the mix” (@RWTH) … in cooperation with Supervised with reinforcement to replace teach-in programming

Setting: To focus servicing activities as far as possible on maintenance and repair, automated detection of defective components has to be sped up through learning systems. The goal is to safe the effort Provisioning for the repairs, avoiding unnecessary controls (from “on-schedule” towards “on-demand”). Algorithm: combining unsupervised and Main- Operations supervised learning for root cause analysis, tenance based on monitoring the condition of the vehicles specifically during operation

Can we improve finding the root cause behind a perturbance, thus improving the maintenance planning? – YES, WE CAN! (2019) Back to  3 shades of AI The magic triangle of Algorithms, 5G and Quantum Computing

TOMORROW . How does connectivity influence AI? . If AI becomes more and more “networked intelligence” – how to connect to (myriads of) devices? . How to enable services in real-time? How to decrease time to insights, decisions and execution? . Which applications become possible Connectivity-driven or improve significantly? phase, . What are paradigm shifts to AI e.g. 5G resulting from enhanced connectivity? CHAPTER 3 AI FURTHER DRIVEN BY CONNECTIVITY

28 The connected world An old vision, now breaking through

“When wireless is Smart perfectly applied, the Building metering whole earth will be converted into a huge brain, which in fact it is, […] and the instruments through which we shall be Smart grid able to do this will be Room amazingly simple compared automation with our present telephone. A man will be able to carry one in his vest pocket.” Smart Nicola Tesla, 1926 environment … and more A short history of mobile radio for voice and data services From 2G to 5G

4G . Designed primarily for data . IP-based protocol (LTE) . True mobile broadband 3G . Lower latency . Designed primarily for voice . Data applications as text, multimedia, internet 2G . First mobile broadband . Designed for digital voice 5G . Introduction of SMS . Enhanced bandwidth . Improved coverage (to 1G) . First digital standard (GSM) . Capable of “True real-time” . CDMA multiplexing applications . Variable devices . With AI capabilities Transmission Rate in Mbit/s Application-specific 5G networks The three primary 5G new radio (NR) use cases defined by the 3GPP Enhanced Mobile Broadband (EMBB) . data-driven use cases requiring high data rates across a wide coverage area . generally: making the mobile internet experience faster and Massive Machine Type more seamless Communications (mMTC / . specifically: from real-time M2M) translation to AR and VR . need to support a very large . higher user mobility – enabling number of devices in a small mobile broadband services in area, which may send data moving vehicles including cars, only sporadically buses, trains and planes . examples: all Internet of Things (IoT) use cases Ultra-Reliable and Low Latency Communications (URLLC) . involves both, industrial and . strict requirements on latency and reliability for mission critical communications production applications as well as networking of . examples: remote surgery, autonomous vehicles, tactile internet, factory automation everyday items such as . applications requiring sub-millisecond latency with error rates that are lower than 1 refrigerators packet loss in 10⁵ packets Separating “brain” and “body” Towards new / renewed steering paradigms of technological systems

Emerging pattern: geometrical distances loosing impact (again)

Distributed intelligence - brain and body not necessarily at the same place but behaving like “one entity“! Brain Body

Combining (in RT) Orchestrating heterogenous data (networked) actors

Coordinating Inferencing seamless training on the edge

Providing computational Possessing a power at scale permanent uplink Platooning of trucks (@RWTH) Network architecture for real-time platooning

Can we steer convoys of trucks fully centralized, thus using the existing computational power in full? – At that time, we could not! (2009)

Setting: Platooning of trucks saves road space, fuel, and enhances security. Distances between vehicles are about 10m at a velocity of 80 km/h (in Japan, even 4m in 2013). Architecture: Without fast communication, the local control With a low latency network (and enhanced algorithms for this multi-agent GPS precision=, the system could be system had to be “onboard”, with steered from a cloud, using the HPC consequences for performance and available, and enhanced scalability, trucks participating (general problem robustness and innovation speed. (2023?) of integrating older systems). Remote and Automated Train Control (@DB) Network requirements for automated train operations

Setting: Autonomous driving is a big topic in AI. For “fenced train environments” as subways, ATO is already reality in a large number of cities. But as for autonomous driving of cars, automated train operations (ATO) Rail capacity and improved energy efficiency in “open environments” depend on by ATO under consideration in Digitale high-speed communication (all 3 Schiene Deutschland (aka “DSD”) and angles, latency, bandwidth and m2m). several other projects, e.g. S-Bahn Hamburg Architecture: Very high reliability in 2021 (2021ff) required, particularly for GoA3 / GoA4. In particular, in critical situations the low latency is necessary for remote Can we run trains highly-automated or even autonomously without making control. intensive changes to the infrastructure? – Today, we cannot! (2019) New models of interaction between human and technology (@DB) AR/VR and hybrid models for customer communication

Samsung (John Godfrey, Setting: Today, customer 11/2018) questions often address “For Samsung, the first answer situative aspects (“are taxis to the question about the available at the south killer app is honestly, that 5G is going to transform the way entrance”)? Employees we interact with technology.” cannot answer these questions, in particular in highly volatile scenarios (e.g. Using more and more cameras everywhere, continuous situative disruptions in operations due information is available over distance “without” delay. Information to weather conditions). will be given based on observation or “on-line data lakes”, and less Architecture: Today, on pre-arranged information bases (2021ff) customer queries are answered on basis of offline Can we answer clients' situated questions, using visual data knowledge or relational over distance? – Today, we cannot! (2019) databases. To summarize: impact of 5G … on AI and on “rest of the world”

Due to the very short latency, 5G reduces the impact of geographical distances.

5G is the basis for the real-time capability of this kind of distributed intelligent systems. Thus, 5G forms an important pillar for a new type of intelligence, based on: 1. Plurality, Mainly due to the enhanced bandwidth, 5G 2. heterogeneity and forms a basis for new interaction models 3. spatial distribution between machine and machine, but also of its components. between human and machine. Back to  3 shades of AI The magic triangle of Algorithms, 5G and Quantum Computing

DAY AFTER TOMORROW . Now, with powerful, yet computationally intensive algorithms at hand and close to zero latency - how can we process the tremendous amount of collected data, preferably in real-time? . How about Moore? How is HPC developing? HPC-driven . How does HPC influence AI? Which phase, models and scenarios become e.g. Quantum Computing possible or improve significantly? . Why is Quantum Computing considered to be “the next big thing”? CHAPTER 4 AI BREAKING GROUNDS VIA NEW TRENDS IN HPC

38 Moore’s Law hitting the ceiling Exponential growth of hardware development stopped?

Moore’s law: The number of transistors on a microchip doubles every 18 month, though the cost of computers is halved. (original quote from Moore, co-founder of Intel, 1965)

A good smartphone 2016 with around 100 billion computing operations per second was almost as fast as the best supercomputer was in the mid-1990s.

About 2016, it became “a common belief” that Moore’s law is coming to its end, due to miniaturization effects (below 5 nm). Hunting for alternatives: Neuromorphic chips Producing a silicon foundation for brain-inspired computation

Goal: . Architecture: emulating the neural structure and operation of the human brain, as well as probabilistic computing, which creates algorithmic approaches to dealing with the uncertainty, ambiguity, and contradiction in the natural world. . Energy efficiency: “Humans run on sandwiches, AI runs on nuclear power plants”

In the European flagship “The Brain Project”, specialized computer architectures are developed, driven by biological paradigms. These architectures are more efficient for certain tasks, but do not follow the “general purpose idea” any longer. Hardware and software become strongly coupled. Thus, experimental changes become more complicated. Hunting for alternatives: AI Bridging Cloud Infrastructure ABCI Super Computer for Artificial Intelligence

Goal: . Architecture: pushing mainly AI, thus optimized for AI applications (e.g., neural networks do not very much depend on high precision); focused on low precision floating rather than Linpack/Lapack performance, GPU NVIDIA based – higher degree of parallel computing . Energy efficiency: “Humans run on sandwiches, AI runs on nuclear power plants” . In Operation: since 2018 A new hope: Quantum Computing (1) The strongest candidate

Goal: . Architecture: The basic properties of quantum computing are superposition, entanglement, and interference. Superposition is the ability of a quantum system to be in multiple states simultaneously.

IBM Q System One (presented 2019), next: 53-qubit quantum computer A new hope: Quantum Computing (2) Solving the trade-off between enhanced computational power and enhanced energy consumption

Enhancing complexity: Quantum computing comes Reducing energy: A quantum processor has to be into play for all or specific types of simulating of complex isolated from its surroundings. This is done by shielding it systems. Today, models for weather, health, chemical and operating it at extremely low temperatures - about 15 models, transportation flows etc. have to be simplified. millikelvin, colder than interstellar space. Hence, the Often, only partitions of the scenario are computed at a processor is superconducting, which means that it can time – thus ignoring their interdependencies and finding conduct electricity with virtually no resistance. This only a set of local optima instead of global ones. processor uses almost no energy and generates almost no heat, so that the amount of energy it consumes is just a fraction of a classical computer’s. Examples for QC applications (@world) Extending the limits

Quantum Healthcare (IBM showcase): When developing new drugs, the test process takes an average of 12 years. Quantum The VW example: The VW engineers want to computers could shorten this period of avoid traffic jams. Without quantum computing, time significantly and, through it would be possible to calculate where intelligent pattern recognition in increased traffic is expected in real-time, which medical data, also identify treatment Google Maps can already show today - but only options that were hidden from classic with the performance of a quantum computer, computers. This would also according to VW, does it seem possible to significantly reduce the costs for the calculate when for each individual car in real- development of new medications and time it is better to turn right or left to prevent would therefore be a hope for patients traffic jams in advance. (since 2018, based on with diseases that are little or not QC D-Wave, published 2019) researched due to their rarity. Examples for QC applications (@RWTH) Integrative simulation over all magnitudes

Can we build a chain of different simulations on different scales that are semantically combined? – Today, we probably cannot! (2018)

Setting: Back at RWTH Aachen, my colleagues in the CoE for

Macroscopic production wanted a “shared model” (of whatever) to work together on. However, they were constantly fighting about “scaling”. Meaning, Mesoscopic which dimension and accuracy to use for descriptions, models and simulations. The material science guys were interested in the Microscopic molecular structure, whereas the people from construction technology wanted different macroscopic dimensions, whereas … Atomic

For an integrated perspective, seamless simulations are needed. Quantum Computing allows the simulation of atoms on the large scale, skipping the reduction of general heuristics in the simulation towards generic applicability (2023?). Examples for QC applications (@DB) The digital twin of the Deutsche Bahn

Setting: Today, only selected aspects of the railway ecosystem can be simulated due to the lack of sufficient DB Simulator HPC power. However, improving the system means mainly integrating the interdependencies of the different components as the failures result form complexity.

Using enhanced HPC power, e.g. QC, we could probably simulate the whole Deutsche Bahn ecosystem, thus Moving from traditional singular aspects over current allowing for new ways of experiment- scaling to interoperable multi-element environments driven understanding the system, from there re-planning, adapting to changes in real-time, understanding the physics Can we create a digital twin of the “complete DB” and run it in a kind of a behind, … (2023ff) simulation environment? – Today, we cannot! (2020) CHAPTER 5 OUTRO

47 Berlin – for the AI community the “ecosystem to be” With BUA right at the heart of it

Broad research StartUps community

Politics and Society BUA Industry

Excellence in Science Wide interdisciplinary scope

Large number of young talents Attractive to brightest minds Beyond scientific excellence: Strengthening German industry and face social responsibility

. Interdisciplinary research, allowing . AI for robust and efficient contributions from software to infrastructures (transport, energy, hardware and everything in between information networks…) . Provide solutions for the society’s . 5G for the core industries of Germany challenges and Europe . Create a German ecosystem for AI in . Speed up the role-out of 5G networks Germany’s capital! . Reinforcement learning in Germany (and Europe) via . Explainable and interpretable AI infrastructure sharing HPC-driven . HMI for AI development in phase, interdisciplinary research communities Connectivity-driven e.g. Quantum Computing phase, Algorithm-driven e.g. 5G phase, e.g. deep learning Prof. Dr. Sabina Jeschke Executive Board DB AG RWTH Aachen University & TU Berlin

Olivier Pfeiffer Think Tank Digitalization & Technology DB AG

Dr. Thomas Thiele Program Manager House of AI DB AG

Jan-Christoph Jähne House of AI Thanks! DB AG //-- Questions? --// Katja Halbmeier Think Tank Digitalization & Technology DB AG Back Up CV

51 CV Prof. Dr. rer. nat. Sabina Jeschke

1968 Born in Kungälv/Sweden

1991 – 1997 Studies of Physics, Mathematics, Computer Sciences, TU Berlin 1994 NASA , Moffett Field, CA/USA 10/1994 Fellowship „Studienstiftung des Deutschen Volkes“ 1997 Diploma Physics

1997 – 2000 Research Fellow , TU Berlin, Institute for Mathematics 2000 – 2001 Lecturer, Georgia Institute of Technology, GA/USA 2001 – 2004 Ph.D. (D r. rer. nat.), TU Berlin, in the field of Computer Sciences; Set-up and leadership of the Multimedia-Center at the TU Berlin

2005 – 2007 Juniorprofessor “New Media in Mathematics & Sciences” & Director of the multimedia-center MuLF, TU Berlin 2007 – 2009 Univ.-Professor, Institute for IT Service Technologies (IITS) & Director of the Computer Center (RUS), Department of Electrical Engineering, 2009 – 2017 Univ.-Professor, Head of the Cybernetics Lab IMA/ZLW & IfU, Department of Mechanical Engineering, RWTH Aachen University

2012 – 2018 Chairwoman VDI Aachen 2014 VisitingProfessor at Hong Kong University of Science and Technology HKUST, Hong Kong since 05/2015 Supervisory Board of Körber AG, Hamburg 2016 VisitingProfessor at Singapore University of Technology and Design SUTD - MIT International Design Centre, Singapore 2011 – 2016 Vice Dean of the Department of Mechanical Engineering, RWTH Aachen University 2017 VisitingProfessor at Volvo Cars, Göteborg/Sweden since 11/2017 apl/Honorary Professorship, Department of Mechanical Engineering, RWTH Aachen University since 11/2017 Member of the Management Board Digitalization and Technology, Deutsche Bahn AG since 04/2018 Chair of Supervisory DB System and DB Systemtechnik, Member of Supervisory Board DB Schenker since 10/2018 Honorary Professorship, Department of Economics and Management, TU Berlin

Senator of acatech and Helmholtz