Nächstes Meetup

Munich Datageeks - November Edition
It is time for our next Meetup. Siemens will host for the second time. Format: • 2 talks (each ca. 40 min incl. discussion) • Time for networking + food + drinks before, in between and after the presentations • Talks are held in English • We will be taking photos and/or film footage at the event. These will be used to share news about our meetups, and to publicize upcoming events. • We will provide a list of registered usernames to the sponsor to check at the entry. The line up: Christian Kroiss: Automatic Hyperparameter Optimization for Machine Learning – An Overview Abstract: With the recent rise of machine learning, it has become usual to see complex models like deep neural networks as part of commercial software development projects. Since the algorithms applied in this field often have a fairly large number of parameters (called hyperparameters), one important question is how to find optimal settings for these parameters that achieve high predictive accuracy while using computational resources efficiently. Traditionally, this tuning task often either requires a lot of experience or involves applying naïve approaches like grid search that can be prohibitively expensive for high-dimensional cases. This talk gives an overview of current approaches and technologies for automatic hyperparameter optimization, which aim to take the burden of manual tuning off the user and at the same time try to consume as few computational resources as possible during search. After briefly introducing some theoretical background, I present a short survey of algorithms that are actively used in practice. Finally, I introduce some publicly available software frameworks (e.g. Tune, HyperOpt) and cloud-based services (e.g. Amazon SageMaker) that implement variants of these algorithms and allow using them in real-world ML projects. Bio: Dr. Christian Kroiß works as a machine learning engineer for Siemens in Munich. Before that, he had been a data engineer, an IT consultant, and a research assistant at the LMU München. During his time as a researcher, he had focused on different topics within AI and formal methods, particularly multi-agent systems, logic-based modeling, simulation, and statistical model checking. Sebastian Schwarz: Calculating CS:GO World Rankings Using RStan Abstract: Today there are tournaments for professional CS:GO teams with prize money as high as $1.5m, top players earn six-digit figures a year - eSports is no longer a niche phenomenon. CS:GO is played as team vs. team (5 players each) in rounds on so-called maps. There is no single league, rather there are multiple usually single or double elimination tournaments like in chess or tennis, so there is no "natural" ranking. There is also no established authority that publishes world rankings calculated with a proper statistical methodology. This talk presents my extension to the bayesian dynamic paired comparison model underlying the Glicko-2 rating system implemented from scratch using RStan. This model not only estimates ratings and enables world rankings, but unlike other heuristics based ratings also features confidence intervals, estimation of additional effects (like T side start (dis-)advantage) and can be used for predicting win probabilities and even score differences of upcoming matches. It will be compared to popular rankings (e.g. on HLTV.org) mainly based on heuristics. Experience during implementation and details in R using RStan will be shown. Code for the scraper, analysis, model and presentation will be open-sourced afterwards. Bio: Sebastian is currently working as Data Lab Manager for E.ON after having gathered experience in the telco business and automotive industry. He studied Economics and Statistics at LMU Munich. At this time he also used to be a professional player for Team Alternate aTTaX (mainly Need for Speed) and took part in (inter-)national events. He always liked competing, when studying in eSports later at Kaggle and Hackathons.

Findet gerade statt

Siemens AG Conference Centre

Otto-Hahn-Ring 6 · München

Du findest uns auch auf