Seminar: Transfer Learning and Continual Learning in Computer Vision (5 ECTS)
Winter Semester 2024/2025, TU München
Organiser: Roman Pflugfelder
Contents
Transfer learning is a very important area of machine learning. The need to reuse learned models with images of unknown scenes, or for new tasks is omnipresent in visual learning. Lifelong learning goes a step further here and allows machines, like humans, to continuously adapt the models without any time restrictions.
Prerequisites
This seminar is for Master's students. Attending the course "Introduction to Deep Learning (I2DL) (IN2346)" and attending an introductory course on machine learning in advance is recommended. Good scientific programming, e.g., Python, Pytorch, Tensorflow, Jax and Flux, is not mandatory for the course. However, mathematical skills and an interest in paper reading and machine learning theory are necessary.
Objectives
The seminar offers the student the opportunity 1) to delve deeper into a topic using a concrete approach (scientific work), 2) to further develop their presentation techniques and theoretical skills through a lecture (15 minutes), and 3) as a presenter and as a listener to practice a constructive feedback culture in the subsequent group discussion. During the seminar, all participants will get to know various methods of transfer learning and lifelong learning and will gain the ability to apply and further develop these methods in their scientific work.
Teaching and Learning Method
During the course, participants will research possible topics together with the support of the seminar leader using collaborative internet tools (Padlet, Miro). Two lectures supplement the basics and individual aspects of the research field. The seminar leader provides individual support to the student during the presentation preparation through individual discussions (Zoom, Slack). Through constructive group discussions during the lectures and the joint development of all individual topics in a seminar report, all students gain a broader knowledge of this field of research.
Registration
Assignment to the lab is done via the matching system.
Timeline
03.07.2024: Pre-seminar meeting (4 p.m. - 5 p.m., room 5609.02.014)15.10.2024: Kick-off meeting (4 p.m. - 5 p.m., room 5609.02.014)22.10.2024: Kick-off meeting (4 p.m. - 5 p.m., Zoom)20.11.2024: Lecture 1 (4 p.m. - 5 p.m., room 5606.01.011)11.12.2024: Lecture 2 (4 p.m. - 5 p.m., room Zoom)21.01.2025: Seminar Day 1 (9 a.m. - 12 a.m., room 5606.01.011)- 22.01.2025: Seminar Day 2 (9 a.m. - 12 a.m., Zoom)
- 12.02.2025: Feedback meeting (4 p.m. - 5 p.m., Zoom)
Material and Links
All material will be shared in Slack.
Grading
- Exercise 1: Participation in the Padlet (+ 0.3)
- Exercise 2: Work out and give a presentation: structure, slides, clarity, depth, own understanding (40%)
- Contribution to the discussion (20%)
- Exercise 3: Contribution to the seminar report: structure, clarity, originality, understanding (40%)