Prof. Dr. Christian Nagel

In this module, students learn to use more advanced algorithms of artificial intelligence and their applications on structures, unstructured and temporal data. The basic idea and mathematical backgrounds of neural networks are introduced. Students learn how to train simple neural networks to learn patterns from data for regression and classification tasks. Further, Deep Learning and its most common architectures are introduced, including Convolutions and recurrent connections. Students learn how to effectively train deep learning networks by choosing optimal hyperparameters and how to avoid overfitting. Thus, methods like Regularization and Dropout are explained. The goal of this module is further to introduce unsupervised learning to the students, as well as its application to solve clustering problems. The application of unsupervised learning in combination with neural networks is illustrated by introducing autoencoders. In addition, it is shown how to use unsupervised learning methods to reduce the dimensionality of datasets using feature selection and PCA techniques. After successfully attending this module, students know:

• How to handle structured, unstructured and temporal data
• What a neural network is and how it can be trained using backpropagation
• How to use different optimizers for neural networks
• The most important deep learning architectural layers like convolutions
• How to effectively train neural networks and to avoid overfitting
• The basic principles of unsupervised learning and their applications to real world problems
• How to used features selection and PCA methods to reduce the dimensionality of datasets
• Different forms of collaborative groups work
• How to gather knowledge and share it within their learning group
• How to summarize and present the most important information of a specific topic

This Master’s seminar studies how to build and improve LLM-based systems when compute is limited, using statistical principles for evaluation, optimization, and decision-making under uncertainty. Rather than training models from scratch, we focus on post-training and inference-time methods and on how to measure what works with reliable, sample-efficient experiments. We cover parameter-efficient adaptation, retrieval augmentation, preference-based optimization, and agentic workflows, with emphasis on benchmarks, uncertainty-aware evaluation, and cost-quality trade-offs.

Topics include:

• Post-training and alignment (instruction tuning, RLHF framing)
• Inference-time reasoning and controllable generation
• Parameter-efficient fine-tuning (PEFT) (e.g., LoRA, DoRA)
• Quantization and efficient adaptation (e.g., QLoRA, recent quantization methods)
• Constrained decoding and logit masking
• Retrieval-Augmented Generation (RAG) as non-parametric memory
• Preference-based evaluation and optimization (e.g., ORPO and ranking-based methods)
• Agentic systems: reasoning–acting loops, tool use, iterative self-improvement, and security