Block-Seminar on Deep Learning

apl. Prof. Olaf Ronneberger (Google DeepMind)

In this seminar you will learn about recent developments in deep learning with a focus on images and videos and their combination with other modalities like language. The surprising emerging capabilities of large language models (like GPT-4) open up new design spaces. Many classic computer vision tasks can be translated into the language domain and can be (partially) solved there. Understanding the current capabilities, the shortcomings and approaches in the language domain will be essential for the future Computer Vision research. So the selected papers this year focus on the key concepts used in todays large language models as well as the approaches to combine computer vision with language.

For each paper there will be one person, who performs a detailed investigation of a research paper and its background and will give a presentation (time limit is 35-40 minutes). The presentation is followed by a discussion with all participants about the merits and limitations of the respective paper. You will learn to read and understand contemporary research papers, to give a good oral presentation, to ask questions, and to openly discuss a research problem. The maximum number of students that can participate in the seminar is 10.

The introduction meeting (together with Thomas Brox's seminar) will be in person, while the mid semester meeting will be online. The block seminar itself will be in person to give you the chance to practise your real-world presentation skills and to have more lively discussions

Contact person: Rajat Sahay

Blockseminar: (2 SWS) In person. Date: Monday + Tuesday, 29. + 30. September 2025, 9:30 - 15:30 in person Room:(tba) Beginning: If you want to participate, attend the mandatory introduction meeting (Will be held jointly with Seminar on Current Works in Computer Vison) on April, 23., 14:00, register in HisInOne, and submit your paper preferences before April, 28.. Mid-Semester Lecture: (tba, via video conference) Introduction to Generative models by apl. Prof. Olaf Ronneberger (Google DeepMind) Recommended semester: 6 (Bachelor), any (Master) Requirements: Background in computer vision Remarks: This course is offered to both Bachelor and Master students. The language of this course is English. All presentations must be given in English. Topics will be assigned for both seminars via a preference voting. If there are more interested students than places, first priority will be given to students who attended the intrdocution meeting. Afterwards, we follow the assignments of the HisInOne system. We want to avoid that people grab a topic and then jump off during the semester. Please have a coarse look at all available papers to make an informed decision before you commit. If you don't attend the meeting (or not send a paper preference) but choose this seminar together with only other overbooked seminars in HisInOne, you may end up without a seminar place this semester. Students who just need to attend (failed SL from previous semester), need not send a preference for a paper, but just reply with "SL only".

---

Material

from Thomas Brox's seminar:

• Giving a good presentation
• Proper scientific behavior
• Powerpoint template for your presentation (optional)

Papers:

Please send your answers to the questions to the corresponding advisor before the seminar day.

The seminar has space for 10 students

No	Paper title and link	Comments
B1	Back to Basics: Revisiting REINFORCE Style Optimization for Learning from Human Feedback in LLMs	This is the RL approach that actually works for LLMs in practice.
B2	DeepSeek-V3 Technical Report	State of the art Open Source LLM. We will select a part of the paper for presentation.
B3	DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning	Open Source reasoning LLM With a performance comparable to OpenAI-o1-1217
B4	Qwen2.5-VL Technical Report	Another State-of-the-art Open Source LLM with strong vision capabilioties
B5	Gemma 3 Technical Report	Another strong State-of-the-art open source model
B6	SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training	RL helps to generalize
B7	Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?	RL is the go-to technique to train the latest resoning models. Maybe it is not as transformative as people think?
B8	Intuitive physics understanding emerges from self-supervised pretraining on natural videos	Learning physics from videos only
B9	SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features	A typical image encoder for LLM's
B10	Scaling Laws for Native Multimodal Models	Early or late fusion of the image encoder into the LLM