WiSe 24/25: Seminar/Proseminar: Large Language Models
Tim Landgraf
Kommentar
This seminar provides an exploration of large language models (LLMs), covering both foundational concepts and the latest advancements in the field. Participants will gain a comprehensive understanding of the architecture, training, and applications of LLMs, based on seminal research papers. The course will be organised as a journal club: students present individual papers, which are then discussed in the group to make sure we all get the ideas presented.
### Potential Topics
- Neural networks and deep learning basics
- Sequence modeling and RNNs (Recurrent Neural Networks)
- Vaswani et al.'s "Attention is All You Need" paper
- Self-attention mechanism
- Multi-head attention and positional encoding
- GPT-1: Radford et al.'s pioneering work
- GPT-2: Scaling and implications
- GPT-3: Architectural advancements and few-shot learning
- BERT (Bidirectional Encoder Representations from Transformers)
- T5 (Text-To-Text Transfer Transformer)
- DistilBERT and efficiency improvements
- Mamba:l and other SSMs: Design principles and performance
- Flash Attention et al: Improving efficiency and scalability
- Training regimes and resource requirements
- Fine-tuning and transfer learning
- Emergence of new capabilities
Schließen16 Termine
Regelmäßige Termine der Lehrveranstaltung