Practical Course: Creation of Deep Learning Methods (10 ECTS)
Winter Semester 2023/24, TU Munich
This is the winter semester 2023/2024 course. For the summer semester 2024 course, see here.
Please send applications (including learning goals, programming skills description, code, all grade transcripts (small PDF file size) - see preliminary-meeting slides) to create-dl[at]vision.in.tum.de
Organizers: Dr. Vladimir Golkov, Lu Sang, Linus Härenstam-Nielsen, Qadeer Khan, Prof. Dr. Daniel Cremers
Preliminary meeting (not obligatory): 5 July 2023, 2pm online: https://bbb.in.tum.de/vla-493-ke7
Slides from an earlier semester's preliminary meeting are available here.
Sending an email to create-dl[at]vision.in.tum.de with sufficient info about yourself (learning goals, programming skills description, code, all grade transcripts (small file size)) until 20 July (ideally earlier) is crucial for matching success. Details about the matching system can be found here and here.
If you ask for a spot after the matching phase, but do not hear from us soon, it means that we cannot offer you a spot.
Content
Using deep learning to solve real problems often requires the creation of novel appropriate deep learning methods, rather than just out-of-the-box usage of existing architectures. In this practical course, students will choose real open problems and learn how to analyze them, how to identify the requirements that a deep learning method should fulfill, and how to create novel deep learning methods that fulfill these requirements.
Some of the projects that can be chosen also include the analysis of design principles of existing methods, and subsequent usage of these design principles to create new methods.
If you want to propose an own project instead of choosing from the projects that we will offer, please discuss with us before 10 July 2023. Use the email subject "PROJECT PROPOSAL" and the aforementioned email address.
Prerequisites
Good programming skills. Eagerness to acquire and deepen knowledge about how to solve complex problems with machine learning. Passion for mathematics. Knowledge of Python and array programming in NumPy (or Matlab or similar) is recommended. Having good soft skills (or the willingness to acquire them quickly) and using them is a prerequisite, particularly good and strategic communication skills, ability to identify what is unclear, to formulate questions precisely.
Course Structure
The students will work individually and in groups on practical deep learning projects. At the end of the project, each student or group will present their project with a following Q&A session. There will be no additional written or oral exam. Both the theoretical and practical part of the project will be considered in the final grading.
Introductory Lectures
The students will receive recordings of introductory lectures as early as they want.
Lecture 1: Machine Learning; Artificial Neural Networks; Convolutional Neural Networks; Q&A about Deep Learning
Lecture 2: Recap; Network Architecture Design; Q&A about Deep Learning
Lecture 3: Recap; Network Training; Understanding and Visualizing; Evaluating; Q&A about Deep Learning
Note: ECTS credits are the measure of workload. So-called semester weekly hours (Semesterwochenstunden, SWS) are NOT a measure of project work time, but merely of classroom time. You will receive remote access to GPUs and are free to work remotely.
Literature
- Christopher M. Bishop. "Pattern Recognition and Machine Learning", Springer, 2006 (Skim the Chapters 1, 2, 5.)
- Good current papers