A WHOLE-DAY MEETUP OF THE MUNICH DATAGEEKS

Abstract: Reinforcement Learning to Run ... Fast

Learning to understand the sensory world is an essential prerequisite for artificial intelligence, but learning how to take actions based on that understanding is the real challenge. Adaptable approach to such sequential decision making, the reinforcement learning (RL), has been applied to various academic and industrial problems by the scientist at NNAISENSE, showing its true long-term commercial potential. This talk will, among other RL industrial applications, introduce the approach that lead NNAISENSE to winning the first place in learning to run competition at NIPS'2017, where a biomechanically realistic model of the human lower musculoskeletal system was controlled to run as fast as possible in 10 seconds of simulated time.

Bio:

Jan Koutnik is a co-founder and director of intelligent automation at NNAISENSE in Lugano, Switzerland. He received his Ph.D. in computer science from the Czech Technical University in Prague in 2008, where he also served as assistant professor in computer science. In 2009 he joined The Swiss AI Lab IDSIA as machine learning researcher. His research was focused on artificial neural networks, recurrent neural networks, evolutionary algorithms and deep-learning applied to reinforcement learning, control problems, image classification, handwriting and speech recognition. In 2014 he co-founded NNAISENSE, which he joined full time in 2016. He was the team leader at learning to park demonstration, shown, together with Audi AG, at NIPS'2016 and learning to run competition, that NNAISENSE won at NIPS'2017.

Abstract: Defense against the dark arts – real-time network intrusion detection

With the democratization of the Internet and the increasing number of connected devices, the number of malicious users intending to cause damage to networks and computer systems is constantly on the rise. Not even companies like Quora, Google and Uber are safe – hackers around the world are getting more sophisticated, but does corporate cybersecurity infrastructure always keep up? A lot of different mechanisms have been developed with the goal of defending networks against attacks and intrusions. The term anomaly-based network intrusion detection refers to finding unusual patterns in network traffic that do not conform to normal activity. The aim of this talk is to discuss how to identify such anomalies from a machine learning perspective.

Bio:

Alisa Krstova is a software engineer currently working at the intersection between photogrammetry, 3D modelling and artificial intelligence. She obtained her Bachelor degree in Computer Science in 2018 from the Faculty of Computer Science and Engineering in Skopje, Macedonia, where she also worked as a part-time teaching assistant. She has been involved in several research projects and scientific publications focusing on image recognition, natural language processing and data mining. Ambivert, curious, she believes that with a cup of good coffee anything is possible.

Abstract: Facing a billion notes

Music can be seen and analyzed as a thing-in-itself, but it is arguably its capacity to make us feel, to project emotions and to be played with that is ultimately interesting. Today it is relatively easy to get hold of and analyze lots of musical artifacts that make up the body of our musical culture, but detailed data about how we interact with music second by second are much harder to come by. In the first part of this talk I describe the process and results of analyzing very large music data sets (hundreds of thousands of hours of music and millions of score sheets), and in the second we will explore an interactive learning approach that puts our personal interaction with music in the center. A special interactive live demo for every audience member will illustrate the presented concepts.

Bio:

Vladimir Viro is the founder of Peachnote, a Munich-based company developing artificial music intelligence technologies and applications. Having received both musical and scientific training, Vladimir combines his love of music and joy of solving technical and business challenges in creating systems that make enjoyment and appreciation of classical music more accessible. His work is reaching millions musicians and amateurs around the world.

Vladimir is also serving as the CTO of the Karajan Institute. He is participating in several international research consortia. He is actively involved in supporting the music information retrieval
education activities at Munich’s universities. He has wide-ranging research interests with the core expertise centered around efficient algorithms, data mining, distributed systems and music information retrieval.

Abstract: Challenges, Limitations, and the Future of Machine and Deep Learning

Current approaches from machine and deep learning in Artificial Intelligence are ill-equipped to deal with highly intelligent abilities such as abstraction, logical inference, and model representation and generation (explicability or interpretability). Judea Pearl, a founder of AI calls machine and deep learning 'sophisticated curve-fitting techniques', and Marvin Minsky, another founder of AI, has proposed to rather study and apply the theory of algorithmic probability to make real progress in AI. We will explore their arguments and possible directions for future research directions in artificial general intelligence despite the success of traditional old-fashioned AI (including deep learning). Current approaches in machine and deep learning devote most of the efforts and resources to dealing with data gathering, cleansing and (over-)fitting but little is invested in producing generative models able to make contributions to causation in areas of science. I will be presenting some of our latest papers in the journals of Nature Communications, and Nature Machine Intelligence, introducing novel methods for causal deconvolution. The techniques are at the core of the new area of Algorithmic Information Dynamics, a field that combines computability and complexity theories, and optimal inference to approach AI bottom-up to meet the current top-down approach based on correlation and regression (as claimed by Judea Pearl).

Bio:

Dr. Hector Zenil co-leads the Algorithmic Dynamics Lab at the Karolinska Institute in Stockholm, Sweden; is co-founder of Oxford Immune Algorithmics; and is the author of around 100 peer-reviewed articles. His website is found at https://www.hectorzenil.net/ and video produced by Nature made to explain one of his latest papers: https://www.youtube.com/watch?v=rkmz7DAA-t8

Abstract: What is intelligence and why the hell are we so obsessed to create an artificial version of ourselves

It seems like the whole world aims to develop artificial intelligence, although nobody really understands what intelligence actually is. How can we try to develop something that we don’t understand and that might not even exist? This uncertainty might be one reason why nearly every discussion on the capabilities of artificial intelligence rapidly converts from a technical into a philosophical argumentation. In this talk we will have a technical and philosophical view on what we commonly understand as intelligence. Elaborating more on the question what intelligence actually means, quickly leads to a conclusion that what we currently understand as artificial intelligence might just be limited to be a replication of our specific human behavior. That again raises the question if machine intelligence is just an imitation and simulation of human intelligence without being able to develop its own consciousness, or if human intelligence is just a result of well-formed and greatly evolved computation power.

Bio:
Dr. Torsten Schön is working on machine learning problems for roughly 10 years where he faced many different problems of different domains. He worked in research, for different startups and in big enterprises. Since three and a half years, he is working with deep learning and at the same time he became father of two kids. Being able to compare human and artificial intelligence made him to question what intelligence actually means and why everybody aims to develop a true artificial intelligence. Further, he is well known for bad jokes and he loves to write 5 sentence biographies about himself.

Abstract: Wearable-based Severity Detection in the Context of Parkinson's Disease Using Deep Learning Techniques

One major challenge in the medication of patients with Parkinson’s disease is that the severity of the disease, reflected in the patients’ motor state, can not be measured using accessible biomarkers. In this thesis we therefore develop and examine statistical models to detect the motor state of those patients based on sensor data from a wearable device on a minute level. We find that deep learning models consistently outperform classical machine learning models applied on hand-crafted features in this time series classification task. Furthermore, our results suggest that treating this problem as a regression instead of an ordered regression or a classification task works best. For consistent model evaluation and training, we adapt the Leave-one-subject-out validation scheme to the training of deep learning models. We also employ a class-weighting scheme to successfully handle the problem of high class imbalances in this domain. In addition, we propose a customized performance measure that reflects the requirements of the involved medical staff on the model. To mitigate the problem of limited availability of high quality training data, we successfully employ a transfer learning technique which helps to improve the final model performance substantially. Since trust in the model’s predictions plays a crucial role in the adoption of those techniques, we additionally extend our model to quantify prediction uncertainty. Our results suggest that deep learning techniques offer a high potential to automatically detect motor states of patients with Parkinson’s disease with great success.

Bio:
tbd.

Abstract: DoDiDa (Doing Dilbert Data)

We love Dilbert. You love Dilbert. How could we not?

But how has this awesome comic strip changed over time? Who is most talkative? What buzzwords were ridiculed and disappeared? Which budgets has the pointy-haired manager denied? And how do you find out out while keeping a full-time job (hint: using sneaky NLP)?

Buckle up for a healthy dose of NLP, graph databases, document analytics, and of course lots of comics in this less-than-totally-serious talk.

Bio:
Rafael is the co-founder and CTO of Curiosity, a deep-tech AI company based in Munich, Germany. Before founding Curiosity, Rafael spent eight years as a researcher in the corporate technology office from Airbus, on topics related to applied mathematics, trajectory optimisation and AI. When not optimizing code, Rafael likes to spend his time skiing or cooking, and can be found on twitter as @theolivenbaum.

Abstract: ML pipelines quality assurance in production

This talk discusses questions such as: How not to fail with your 99%-accuracy model in production? Which metrics should be checked and when? How do you control live predictions? What to do if your model accuracy is degrading? Using some live examples, we will cover main quality assurance steps that can be applied alongside model development and deployment, and we present the practices of unit testing and results monitoring for production.

Bio:
Alex Tselikov works as a lead data scientist for KI labs in Munich where he is responsible for building machine learning and data engineering products. Before that, he had been a senior data scientist at Veon (one of the biggest telco in Russia with 50 mlns subscribers) where he had focused on applying AI algorithms for credit scoring, customer churn analysis, fintech, chat-bots with neural networks and natural language processing. He had also been responsible for building and deploying large-scale machine learning pipelines into production environments. He holds his Ph.D. in data analysis from MSIU (Russia) and kaggle expert rank.

Abstract: Deep Learning for self-driving cars: Potentials and limits

Self-driving cars are one of the biggest fields of application for Deep Learning. While Deep Learning has substantially increased their capabilities, it also brought new challenges in terms of system architecture and interpretability. This session aims to explore the application of state-of-the-art methods as well as to look at some of these new challenges.

Bio:
tbd.