Seminar for computer science master students (IN2107).

Update (Oct 8): initial meeting shifted to October 29, 4.15pm-5.45pm, room 02.09.023

Modern deep learning methods have shown remarkable success in various domains of machine learning. However, our theoretical understanding of such methods remains rather shallow. Recently, novel ideas to better understand aspects of optimization and generalization in these models have emerged that require rethinking classical concepts in these domains. A prominent example is the implicit bias of optimization methods in an overparameterized regime. We will discuss such ideas from the computational and statistical viewpoint.

Seminar material and the schedule can be accessed here.

• Organizer: Thomas Frerix
• Email: DL-seminar@vision.in.tum.de

• Date: Wednesday, July 17, 2019
• Time: 4.15pm
• Location: 02.09.023

• Date: Tuesday, October 29, 2019
• Time: 4.15pm-5.45pm
• Location: 02.09.023

• Registration is done via the TUM matching platform
• Attendance in the seminar sessions is mandatory
• Preference is given to students with with a pertinent background in relevant fields of applied mathematics (Linear Algebra, Analysis, Numerical Optimization, Probability/Stats) and prior exposure to deep learning (e.g. Introduction to Deep Learning).
• Important: When registering at the TUM matching platform, please also send an email to DL-seminar@vision.in.tum.de in which you write your full name, matriculation number and where you state and prove prior exposure to these topics, e.g., with a transcript of records, prior course projects etc..

• We will hold a one-week block seminar, December 2 - December 5, 6-8pm. Room: 02.09.023
• This way the seminar does not conflict with the exam period and the evening times hopefully avoid conflicts with other lectures.
• All talks are held in English.

Final reports should be written in LaTeX. A template will be provided.