Seminar: Self-Supervised Learning
Prof. Dr. Abhinav Valada1
Co-Organizers:
Juana Valeria Hurtado2 Dr. Daniele Cattaneo3 Dr. Tim Welschehold4 Daniel Honerkamp5
Self-supervised learning is an exciting new paradigm that aims to learn representations from data without explicit supervision. Self-supervised learning has emerged as an alternative to classical supervised learning methods that strongly rely on large labeled datasets for training. Creating datasets with sufficient annotated examples is a challenging task, and the process of labeling data is often arduous, expensive, and sometimes even infeasible. Self-supervised learning mitigates this problem by first learning a pretext task exploiting some property of the data and then using the learned semantically rich representations for fine-tuning on the target task. Here's a recent post that explains this in more detail6. In this seminar, we will study a selection of state-of-the-art works that propose self-supervised approaches for perception, state-estimation, and control.
Details
Seminar: | Online via Zoom7, Meeting ID: 939 8825 4440, Password: !SSL2020! Places: Max. 12 students |
Beginning: | Friday, May 15, 2020, 14.00-16.00 Introduction via Zoom7 and allocation of seminar topics via email Introductory lecture slides can be downloaded here8 |
Requirements: | Basic knowledge of deep learning or reinforcement learning. |
Remarks: | Due to the Corona crisis, the entire seminar will be held online. Topics will be assigned for the seminar via a preference voting. If there are more interested students than places, places will be assigned based on priority suggestions of the HisInOne system and motivation (tested by asking for a short summary of the preferred paper). The date of registration is irrelevant. In particular, we want to avoid that students grab a topic and then leave the seminar. Please have a coarse look at all available papers to make an informed decision before you commit. |
Procedure
- Enroll through HISinOne9, the course number is 11LE13-7314-M. Registration period for the seminars in HisInOne are from 11.05.2020 00:00 to 20.05.2020 12:00.
- Attend the introductory session on Friday, May 15, 2020, 14.00-16.00, via Zoom7, Meeting ID: 939 8825 4440, Password: !SSL2020!
- Students should select three papers out of the list in preference order and fill the information in this form10
- Places will be assigned based on priority suggestions of HisInOne and motivation of the student by May 24, 2020.
- Students are requested to prepare a 20 minutes talk, write an abstract and a summary.
- The Seminar will be held as a virtual "Blockseminar" in the last week of July.
- The details of the presentation and the slides should be discussed with the supervisor two weeks before the presentation.
- The abstract should be two pages long and is due June 29, 2020. Please use this template11.
- The summary is due two weeks after the presentation and should be seven pages long at maximum (latex, a4wide, 11pt) not counting the bibliography and figures. Significantly longer summaries will not be accepted.
- The final grade is based on the oral presentation, the written abstract, the summary, and participation in the blockseminar.
Topics
- SuperPoint: Self-Supervised Interest Point Detection and Description12
Supervisor:Dr. Daniele Cattaneo3 - GANVO: Unsupervised Deep Monocular Visual Odometry and Depth Estimation with Generative Adversarial Networks13
Supervisor:Dr. Daniele Cattaneo3 - SelFlow: Self-Supervised Learning of Optical Flow14
Supervisor:Dr. Daniele Cattaneo3 - A Simple Framework for Contrastive Learning of Visual Representations15
Supervisor:Daniel Honerkamp5 - Grasp2Vec: Learning Object Representations from Self-Supervised Grasping16
Supervisor:Daniel Honerkamp5 - Visual Reinforcement Learning with Imagined Goals17
Supervisor:Daniel Honerkamp5 - Improving Semantic Segmentation through Spatio-Temporal Consistency Learned from Videos 18
Supervisor:Juana Valeria Hurtado2 - VideoBERT: A Joint Model for Video and Language Representation Learning 19
Supervisor:Juana Valeria Hurtado2 - Self-Supervised Scene De-occlusion 20
Supervisor:Juana Valeria Hurtado2 - Imitation from Observation: Learning to Imitate Behaviors from Raw Video via Context Translation21
Supervisor:Dr. Tim Welschehold4 - Unsupervised Perceptual Rewards for Imitation Learning22
Supervisor:Dr. Tim Welschehold4 - Time-Contrastive Networks: Self-Supervised Learning from Video23
Supervisor:Dr. Tim Welschehold4
What to put into the summary?
Some questions that should be discussed are the following:
- What is the paper's main contribution and why is it important?
- How does it relate to other techniques in the literature?
- What are strong and what are weak points about the paper?
- What would be interesting follow-up work? Any possible improvements in the methods? Any further interesting applications?