Veranstaltung | Wochentag | Termin | Ort |
042531 (Übung: 042532) | Montag (Übung: Mittwoch) | 10.15 - 12.00 | Otto-Hahn-Str. 12 - R 1.056 Campus Nord |
This course focuses on optimization techniques to find solutions of large-scale problems that typically appear in statistical learning / data analysis tasks with big data. Topics would include widely adopted methods in modern research literature such as projected gradient methods, accelerated first order algorithms, conjugate gradient methods, quasi-Newton methods, block coordinate descent, proximal point methods, stochastic sub-gradient algorithms, alternating direction method of multipliers, and semi-definite programming. Efficiency of these methods in terms of convergence rate and overall time complexity will be discussed, so that one can see how to choose a suitable method for his/her own research problems. Separable (approximate) reformulations of optimization problems will also be discussed that are suitable for parallel computation compared to their original formulation. Discussions in the course will be problem- oriented, so that methods will be presented in the context of specific problem instances whenever possible. Homework assignments will be given to provide background knowledge for the course or to check understanding of techniques.
The aim of this course is to provide students with understanding of modern optimization techniques suitable for large-scale/big-data problems, so that students see how they can choose, apply, and/or modify appropriate methods for their own research problems.
This lecture will be based on recent publications that will be assigned as reading homeworks. No textbook is required. Another lecture of mine will be helpful if you need background knowledge:
Some of my lecture materials are from:Due to the nature of practices, only some parts of Ubung materials will be available here. We'll use the Julia language.
*** !!! To register for the final exam, bring this form filled-in by the last lecture date (21.07) to get my signature !!! ***
Final exam registration form
The basic idea is that you pick a paper from a suggested list or of your own choice, and implement the algorithm in the paper. You give two presentations.
The list will be provided soon and keep growing. Discussion with the lecturer is highly recommended for your choice (you must send me an email about it at least). More than one person can work on the same paper, and you can focus on a part of a paper if you have a good reason. Finding more recent papers of similar topics are welcomed. Choose one you can implement around 2 -- 3 weeks. Implementation in the Julia language is strongly recommended. Some papers may accessible only from the university.
>> Methodology-focused (this is the preferred way for a mini-project):
>> Application-centric (interesting application + basic method is possible. Finding newer/better papers is recommended)
Data sets that might be useful for your project.