Machine learning algorithms find its way into an increasing number of (safety-critical) application domains but their quality is rarely assessed in a systematic way. The focus of this module are quality criteria for machine learning algorithms, in particular of deep neural networks, ranging from performance evaluation over explainability/interpretability (XAI), robustness (adversarial robustness, robustness against input perturbations), uncertainty quantification, distribution shift, domain adaptation, fairness/bias to privacy. The methods are introduced theoretically in the lecture and implemented/applied practically in the exercises. Prerequisites are basic theoretical knowledge in machine learning, programming skills in Python and ideally practical knowledge in training neural networks.

This lecture will provide a general introduction to modern deep learning methods with a particular emphasis on practical applicability. At the same time, the course will provide an introduction to the popular PyTorch Deep Learning framework while requiring only basic programming skills in Python. The course will cover a range of common machine learning tasks across different data modalities ranging from tabular data over Computer Vision (image classification, image segmentation) to time series and natural language processing. It will cover the most important model architectures in these domains ranging from convolutional neural networks over recurrent neural networks to transformers. The lecture will be accompanied by a tutorial class where students are supposed to acquire hands-on experience in working with PyTorch and are supposed to acquire the skills to apply Deep Learning methods in their respective fields of study.

This lecture provides insights into state-of-the-art deep learning method for the analysis of medical data. We cover a broad spectrum of data modalities and applications and try to discuss both methodological knowledge as well as the necessary medical background knowledge. In particular, we cover physiological time series (ECG, EEG), medical imaging (histopathology, CXR, CT/MRI), audio data (e.g. from digital stethoscopes), electronic health records, clinical text data as well as multimodal combinations of these data types. The students are supposed to work towards a final project of their choice during the second half of the course.

This lecture encompasses two primary subjects: self-supervised learning and modern generative models. In the first part, we will examine the fundamental design principles (contrastive versus non-contrastive) underlying self-supervised learning algorithms. In the second part, we will explore applications of these principles to specific data modalities such as computer vision, natural language processing (including an extensive coverage of large language models) and audio/time series. Finally, the third part will focus on generative models, where we will cover a wide array of models, ranging from autoregressive models, variational autoencoders, and normalizing flows, to generative adversarial networks and (latent) diffusion models.

This seminar is supposed to cover current publications and/or research topics in the domain of machine learning with particular regard to applications in the health domain. This includes topics with a strong methodological focus (such as self-supervised learning, quality criteria for ML algorithms such as interpretability/uncertainty quantification) but also medical application topics.

This seminar will cover different aspects of interpretable/explainable AI (XAI) ranging from inherently interpretable models over perturbation-based methods, such as Shapley values, to gradient-/decomposition-based approaches and their quantitative evaluation. Going beyond conventional single-feature attribution methods, we will also discuss current concept-based attribution methods, ways to assess feature interactions and connections to causality.

This seminar will cover current approaches to improve the label-efficiency of machine learning systems in different application domains such as computer vision, speech and natural language processing. A particular focus will lie on self-supervised learning but we will also cover aspects of self-training and weak supervision.