Seminar: Neural Network Design Patterns in Computer Vision (5 ECTS)
Winter Semester 2023/24, TU München
Organiser: Roman Pflugfelder
Description
Computer vision considers models of neural networks originally invented by the machine learning community. This seminar allows the student to study neural networks in more detail. By the end of the seminar, all participants will understand the principles and the applications of selected models, and they will receive the ability to reuse these models as common design patterns in their future work.
Prerequisites
This seminar is for Master's students. Attending the course "Introduction to Deep Learning (I2DL) (IN2346)" in advance is recommended. Good scientific programming, e.g., Pytorch, Tensorflow, Jax and Flux, is not mandatory. Analytical skills and interest in paper reading and theory are necessary.
Schedule
The participants will collect a pool of successful neural networks, network architectures and network layers. After an individual choice, each student will read the respective paper publication to search for and study related papers. Based on the findings, the student will work out a short presentation (15 minutes), which will be given to the group. The students learn collaboratively about the model and the findings within a group discussion and feedback. After the presentations, the student will individually work out his findings, including written feedback, i.e., text, images, and videos. At the end of the seminar, the participants collect all these contributions into a final seminar report.
Timeline
03.07.2023: Pre-seminar meeting (4 p.m. - 5 p.m., in presence, Room 03.13.010)18.10.2023: Kick-off meeting (4 p.m. - 5 p.m., in presence, Room 02.07.014)Exercise 1: Collect design patterns and the papers (online using [[https://padlet.com|Padlet]])08.11.2023: Paper selection meeting (4 p.m. - 5 p.m., in presence, Room 02.09.023)Exercise 2: Study papers, do literature research, work out the presentation (on demand, individual [[https://tum-conf.zoom.us|Zoom]] meetings)12.12.2023 - 14.12.2023: Discussion groups (9 a.m. - 12 a.m., in presence, Room 02.09.023)Exercise 3: Work out your contribution to the seminar report (on demand, individual, Zoom meetings)29.01.2024: Final meeting (11:30 a.m. - 12:30 p.m., Zoom Meeting)
Material and Links
- Slides of pre-seminar meeting
- Slides of kick-off meeting
- Kick-off meeting Mentimeter PDF
- Slides of pattern-selection meeting
- Padlet collaboration link
- Padlet result PDF
- Padlet participation overview PDF
Seminar Report
Grading
- Participation during Exercise 1. Own selection of a design pattern (10%)
- Presentation: structure, slides, clarity, depth (20%)
- Contribution to the discussion (10%)
- Written contribution to the seminar report: structure, clarity, originality (30%)
- Understanding of the design pattern and scientific background as part of the written contribution to the seminar report and the individual presentation and discussion contributions (30%)
I will consider your own programmed examples in the grading (+10%)