Big Techday 13

With the ubiquitous rise of AI and the wish to really harvest its possibilities, a new challenge arises which is very different from the rather scientific task of designing good machine learning models: productionizing AI. How can we easily and quickly move our models from proof of concept to 24/7 operations? What do we need to do to ensure our AI processes run smoothly and reliably? Do our solutions scale to production data volumes? How can we track and maintain model quality with limited data science resources, and dozens or hundreds of models in use in operational processes? The Data Analytics and Artificial Intelligence Department at Telefonica with the help of TNG has built an open source-based framework to address and answer these questions on an enterprise level. Our telecommunications industry case study discusses the challenge of applying machine learning to an environment far removed from academic data science - the realtime, 24/7 world of mobile network operations.

Software and technology have changed every aspect of the world we live in. At one extreme are the ‘mission critical’ applications - the code that runs our banks, our hospitals, our airports and phone networks. Then there’s the code we all use every day to browse the web, watch movies, create spreadsheets… not quite so critical, but still code that solves problems and delivers services. But what about the code that only exists because somebody wanted to write it? Code created just to make people smile, laugh, maybe even dance? Maybe even code that does nothing at all? Code that was created just to see if it was possible? Join Dylan Beattie - programmer, musician, and creator of the Rockstar programming language - for an entertaining look at the art of code. We’ll look at the origins of programming as an art form, from Conway’s Game of Life to the 1970s demoscene and the earliest Obfuscated C competitions. We’ll learn about esoteric languages and quines - how DO you create a program that prints its own source code? We’ll discover quine relays, code golf and generative art, and we’ll explore the phenomenon of live coding as performance - from the pioneers of electronic music to modern algoraves and live coding platforms like Sonic Pi.

While Artificial Intelligence has been very successful in many areas, such as medical diagnostics, facial and voice recognition, self-driving cars and other forms of 'pattern recognition', its applicability to the area of stock exchange trading - despite the high level of technological developments - is not very common. This talk points out why the classic 'black box' procedures are structurally more difficult to apply in the financial market than in other areas and at the same time shows that hypothesis-based procedures in the algorithmic trading can still create great value. The second part of the talk takes a social perspective on the medium and long-term opportunities and risks of AI applications. First, the thesis of a possible 'superintelligence' is critically examined, according to which an 'intelligence explosion' could occur, in which AI systems improve themselves iteratively and would thus be beyond any human control. This idea is countered by the structural limitations of current AI systems. At the same time, the focus is on the immediate ethical questions that already arise from the use of such systems today. On the one hand, this concerns the question of how we see ourselves as individuals, whose even most personal decisions are being taken more and more by machines. On the other hand, there is also the political question of whether a data-driven world could promote more centralistic policy approaches over the decentralized processes of democratic decision-making.

Chemistry has always been a discipline of trial and error. Due to the complex relationships between the structures and properties of chemical substances, a downright zoo of heuristics and recipes has been developed, whose complexity can easily overwhelm even experts. On the one hand, the application of the right models costs a lot of time and money and, on the other hand, prevents that enough time remains for the crucial creative phases of scientific work. This talk will focus on how complex tasks such as drug design, reaction optimization, and the retrosynthetic decomposition of molecules into commercially available precursors can be automated using AI methods. Thus, the role of humans in these areas is shifting away from trial and error towards evaluation and refinement of the procedures developed by AI. This approach makes it possible to generate new active ingredients with predicted structure-activity relationships as well as ready-made "recipes" for their production in rapid iterations. The truly time-consuming steps can be taken over by the computer, nearly in real time. In particular, we will focus on the subfield of retrosynthesis, whose complexity is particularly striking. Until the emergence of machine learning methods, this area required the mastery of thousands of reaction regulations, their reaction conditions, and possible dangers. We will also present and critically examine our proof of concept for such software, NeUReAc. NeUReAc was born out of an idea that Segler et. al. published in 2018 in a groundbreaking paper but aims to be easy to use and expand.

Helena’s use of GANs is the modern-day equivalent of da Vinci grinding his own pigments —Grady Booch
Helena invites you on this behind-the-scenes tour as she goes over her experiments with AI as a new artistic medium. She starts with general principles of AI Art making process, its roots in the computational generative art landscape and its tools based on deep neural networks. How does AI art reflect visual truth in times when breakthrough research allows to generate diverse images with photographic accuracy, in our increasingly AI-driven world? From that perspective Helena examines how imagery produced by generative models could be similar to human memory - as collages, as compression and distillation of previous events. Going from general to personal, Helena shares her approach to AI art making for which she coined a term - Folk AI Art: though made using state-of-the-art technology, it aims to carry on traditions of folk art - beauty, humor and craft.

The training of artificial intelligence systems can be accelerated by using simulated environments that deliver realistic data for complex problems. Based on the by TNG developed virtual reality game “Blobby VR”, we show the linkage of Reinforcement Learning to a simulated environment for the training of a computer game AI. Reinforcement Learning represents several methods of machine learning. An agent maximizes received recompenses by learning a strategy independently. For this example, we use a simulation to accelerate the learning process of a deep neuronal network. During the presentation, we are going to discuss the basics of Reinforcement Learning and the simulation first. / We are going to start the presentation with the basics of Reinforcement Learning and the simulation. Afterwards, we are going to illustrate how a linkage between those two technologies within the game engine Unity is possible. In doing so, the simulated environment reproduces reality in all relevant aspects and offers the possibility of replacing human gamers with AI-agents. Using Reinforcement Learning, the behavioural traits of a real gamer can be learned. To conclude, we are going to show how our example is transferable to other applications, for example in the production or logistics industries.

Artificial Intelligence can now be used in Olympic Sports to create technical evaluations for athletes in real time and thus replace hours of post-processing of the videos by trainers. In collaboration with the Olympic Training Centre in Heidelberg, the TNG Innovation Hacking Team created a software that automates scientific technology analysis in top-class canoeing. Video recordings once had to be exhaustively watched and individual images had to be analyzed for specific angles by hand. Thanks to our application, this evaluation is now fully automated using several modern techniques from the area of neural networks. In this talk, you will learn about how the software uses the raw video material to determine the angles at which a canoeist sticks his paddle into the water. The speakers will show you how to stabilize the shaky camera images against the waterline using segmentation, how OpenPose finds the hands of the athlete and how a classification model uses transfer learning to differentiate between a paddle above and below the waterline. The TNG-Consultants will also go into details of the technologies and neural networks used.

The BMW Group decided to re-architect and move its on-premises data lake to the Amazon Web Services Cloud. The Cloud Data Hub (CDH) processes and combines anonymized data from vehicle sensors and other sources across the enterprise to make it easily accessible for internal teams creating customer-facing and internal applications, now and in the future. Currently, millions of BMW and Mini vehicles are connected to the CDH via BMW Group’s highly secure backend, processing terabytes of anonymous telemetry data daily. The company uses this data to monitor vehicle health indicators such as check control errors to identify potential issues across vehicle lines ahead of time. This enables the BMW Group to leverage fleet data ingested, collected, and refined from the CDH to better resolve issues, even before they impact customers. TNG has been supporting this project for roughly one year. Join Josef and Max for a whirlwind tour of the CDH’s vision, its challenges and technical solutions.

The evolution of computer programming languages from machine code to high-level languages is generally well understood. But did you know that a similar advancement is taking place in the world of Cloud infrastructure programming right now? "Infrastructure as Code" (IaC) is not even 10 years old and paradigm shifts abound. Most recently, AWS introduced the Cloud Development Kit (CDK), which combines the convenience of CloudFormation with the power of general-purpose programming languages. In this talk we will provide a general overview of the IaC landscape with focus on Terraform and the CDK. Expect live demos, code snippets, and practical examples of how you can use these tools in your own production environments. If you have ever wondered how you could deploy your infrastructure reproducibly, using powerful abstractions, and without drowning in YAML files, this is the talk for you. This talk is aimed at developers who have working familiarity with at least one public cloud service. Some concepts will be reviewed briefly.

Azure Functions are the serverless offering in the Microsoft Azure Cloud. A hugely popular computing paradigm, serverless seeing increased interest in the last several years. Azure Functions offer several very interesting value propositions, such as deployment slots, real-time monitoring and the support of rich HTTP interfaces, setting it apart from AWS Lambda. In this talk, we will take a look at the most important features and use-cases for Azure Functions, as well as some common pitfalls.

With a length of ten metres, a wingspan of 4.30 metres and 400 HP output, they built the world's largest airworthy model of the Concorde. It is powered by four real kerosene turbines, each with an output of 90 kW and is made of 100% carbon fibre. The manufacturers also relied on digital transformation using a modern remote control system by combining the System Core and Power Box, the safest system in the world.

The American billionaire and visionary Elon Musk builds spaceships and electric cars. In 2013 he published a traffic concept for the future. At more than 1000 kilometers per hour, people will speed towards their destination in vacuum tubes, saving energy and time. But not only the concept is revolutionary, the way to get there is, too. Elon Musk regularly organizes competitions in which the best student groups from around the world compete against each other with their specially developed tube vehicles. The TUM Hyperloop group of the Technical University of Munich, four-time winner of the competition and current world record holder with a speed of 482km/h, has been successful from the very beginning. Longtime member Florian Janke presents the philosophy of the Hyperloop concept, the challenges of the competitions and the groundbreaking developments of the team.

Robotics Erlangen: Autonomous soccer robots in the RoboCup Small Size League:
The “Robotics Activities Erlangen e.V.” and later “Robotics Erlangen e.V.” started as a small project at the University of Erlangen-Nuremberg. Since then we have devoted ourselves to the topic of robot soccer and regularly compete in the RoboCup. At this world championship of robots, we compete in the Small Size League (SSL) against international opponents and were able to achieve the vice world championship title in 2017 and 2019, among others. During the games, the robots must act completely autonomously, and humans are no longer allowed to intervene. From above, cameras watch the game and a central AI controls up to 11 robots. In terms of their mechanical and electronic structure, these were developed and built by the team themselves. The firmware, controls and strategy are constantly being developed. In addition to the challenges that arise in the development as a student project, we will also give an insight into our research and the current state of the art. We show why the acceleration of golf balls for passes must be considered, why highlighters can increase the precision and how forgotten code sometimes ensures that nobody shoots at the goal.

After many years as a software developer in various startups, Teja Philipp created for fun the laser cutter - Mr. Beam. The device's surprising Kickstarter success made him an overnight entrepreneur. 6 years later, he reports joy and suffering on the way from techie to managing director, from kit to product and from crowdfunding to VC investment.

Inside look at a design and build process of a floating remotely-controlled 2-meters-long LEGO model of the Glomar Explorer ship, and how it was used to recover a model submarine in a swimming pool.

Note: This is a follow-up project to the Big Techday 2019 talk "A Real-Life Mission Impossible: Raising a Sunken Soviet Submarine"

It is the year 1561. King Erik XIV of Sweden is involved in a war with Denmark and the Hanseatic city of Lübeck. In order to gain supremacy over the maritime trade routes of the Baltic Sea, he orders the construction of Mars, a new type of warship. After only one year in operation the ship sinks and remains missing for over 400 years. In 2011 the ship was rediscovered near the Swedish island of Öland - a sensation in underwater archaeology! Since then, annual expeditions have brought to light new findings. Photogrammetric models of the wreck lying at a depth of 80m and some artefacts were generated from thousands and thousands of photos and made available by the scientists. These contents were optimized by us especially for the mobile Oculus Quest VR glasses and made accessible in an interactive application. Research results and background knowledge about the legendary battleship can be explored virtually by researchers as well as laymen in playful dives: Warship Mars. Digital archaeology - experience it at home!

We'll go through best information security practices and top them up with thoughts on why that has not been enough to keep us protected.

Which framework is suitable for my (enterprise) application in 2020? The talk offers a comparison of aspects relevant to one's decision making when choosing an option. Which product is to be developed? Is the application to be rewritten or migrated? Which developers are available? How important is performance? How important is modular architecture? How long should the application be used? How important is the flexible use of existing libraries? How important is testability?

While Java’s strict type system is a great help for avoiding programming errors, it also takes away some of the flexibility that developers appreciate when using dynamic languages. By using runtime code generation, it is possible to bring back some of this flexibility back to the Java virtual machine. For this reason, runtime code generation is used by many state-of-the-art Java frameworks for implementing POJO-centric APIs but it also opens the door for assembling more modular applications. This presentation offers an introduction to the complex of runtime code generation and explains why non-standard Java runtimes like Android Runtime or Graal’s Substrate VM sometimes struggle to execute existing Java code. Finally, the presentation details strategies of how library developers could work around these limitations by changing their code generation patterns.

In a much-noticed article, the Nobel Laureate Eugene Wigner talked about the “unreasonable effectiveness of mathematics in the natural sciences”. The best way to clearly experience this wondrous effect of mathematics on physics is through the formation of spectral colors of diluted hydrogen gas. For this, no more previous mathematical knowledge than what you have learned in school is necessary, nothing more than division…

We will present a new approach and its implementation for the migration of legacy code, which are based on test-driven development and customer-specific extensions. The code is automatically translated from an outdated source language like PL/I to a modern language like Java. This prevents repetitive and error-prone work during the translation. At the same time, by using customer-specific heuristics, we avoid over-generalizing the generated code and it remains easy to read. In the talk we first deal with how we map general features of the source language using a context-free grammar and then automatically translate them. The integration of customer-specific features and the iterative expansion of the translator (transpiler) also represent an important cornerstone of our approach. We will show this using the example of selected preprocessor directives. We also go into the module test framework for PL/I created especially for this purpose and its use in test-driven translation. Finally, we examine the reliable maintenance of the semantics of the original code and thus pave the way for subsequent refactoring in the target language.

Lessons from nearly 25 years spent working with, for and against Google and what these lessons mean for a future increasingly dominated by monopolists.

Modern work is more complex and interdependent than ever before. In a recent study, Slack explored what workers need to thrive in this new era of business. We found there is a fundamental challenge common to all organisations. And that challenge is, at the simplest level, one of an alignment paradox. Using our own Solution Engineers as an example we want to show how a day in their life looks like. Making people's working lives simpler, more pleasant and more productive by leveraging integrations, workflows and apps in Slack.

Can the intranet be expanded in a user-friendly and mobile manner as an end point for communication, HR services and IoT? How can communication work effectively, quickly and easily using private tools - BUT still be compliant with the GDPR and worker’s council (Betriebsrat) guidelines? And how can top-down information be distributed to ALL employees - whether in production, in the warehouse, at the desk or part-time - in real time? Benedikt Ilg, CEO of Flip, explains how to solve ALL these problems with just one app.

Distributing and sharing files between coworkers and friends is a problem that’s been around for a while. In the last 10 years, services like Dropbox have used cloud infrastructure to help make data sharing smoother. While convenient, these centralized services come with serious drawbacks. In this talk I’ll introduce an open-source, P2P alternative we’ve been working on called Hyperdrive, and why we think it finally offers a good and competitive solution to existing products. I’ll make sure to give a deep technical analysis and have plenty of demos.

The Christmas season, initiated by Black Friday, is the period with the highest turnover for many B2C webshops. Therefore, it is especially important for fast-growing shops to ensure in advance that the application and infrastructure will be able to handle the expected load. We report on our experiences last year in preparing a customer shop for the pre-Christmas season. In particular, we will discuss how the interplay of distributed load tests of the production system and suitable data visualization can be used to identify performance bottlenecks in a web application on the one hand and to increase stakeholder confidence in the entire application setup on the other.

JHipster is the most popular code generator in the Java world with more than 14k GitHub stars, 500+ contributors, 100ths of companies using JHipster and a large ecosystem of modules and plugins. This session will introduce you to the possibilities of JHipster and show you how to build a modern web application with JHipster using Java and JavaScript. We will show the typical application scenarios and build from scratch to a complete web application and run it in the cloud. We will take a look at best practices in dealing with generated code and give an outlook on the future of JHipster.