Machine Learning for the Natural Sciences
This lecture series provides a comprehensive introduction to machine learning techniques and their application to challenges in the natural sciences. Starting with foundational concepts like classification and regression, the course prog
resses through key models including tree-based methods, logistic regression, artificial neural networks, convolutional neural networks (CNNs), and transformers. Students will gain hands-on experience implementing these methods using PyTorch and scikit-learn, culminating in a series of capstone projects designed to apply learned techniques to real-world scientific problems. Topics also include uncertainty quantification (UQ), explainable AI, and an exploration of cutting-edge generative models like variational autoencoders, diffusion models, and large language models (LLMs). This series aims to equip learners with the tools and knowledge to leverage the power of machine learning in their respective fields.
Planned Topics:
- Classification with Decision Trees and logistic regression
- Classification metrics and cross validation
- Classification: confusion matrix, ROC, (bias due to training set selection?)
- from classification to regression: MLE, least squares method, linear models
- regression: ANNs and SGD
- regression with pytorch
- activations functions and classification with pytorch
- convolution and CNNs
- ResNets and UNets
- Transposed Convolutions, Auto-Encoders and Variational Auto-Encoders
- Variational Auto-Encoders and Normalizing Flows
- Diffusion and FlowMatching
- Uncertainty Quantification
- Methods of Explainable ML
- ML with Text, Attention and Transformers
- Large Language Models (LLMs) and chatGPT
Venue
- every week: Friday 4th block (13:00-14:30), ASB/328
- every second week: Wednesday 6th block (16:40-18:10), SE2/201
House keeping details
- This course will be given entirely in English.
- Please use the enrollment to be part of the lecture.
- The course comprises, lectures, labs and a capstone project.
- The lecture comprises 3 SWS and 4.5 ECTS credits if this is needed.
- If you like to pass the lecture, you have to complete a capstone project.
Video Conference Details if required
Join Zoom Meeting
hzdr-de.zoom-x.de/j/89839487704?pwd=QXp3MWc4WVNTRFdBeldhRWVBU0RwUT09
Meeting ID: 898 3948 7704
Passcode: 767523
Important Dates
capstone project handout to students: Dec 1, 2025
poster submission for print: Feb 2, 2026
poster presentation: Feb 6, 2026
Lecturers
Peter Steinbach: Peter received his PhD in Particle Physics in 2012 from TU Dresden (Germany) for an experimental study of LHC data using the ATLAS experiment to reduce background contributions to Higgs Particle searches. He continued to industry (Scionics Computer Innovation GmbH) as a HPC support and software engineer helping scientists push the limits of their applications in a service oriented group at the Max Planck Institute for Molecular Biology and Genetics. In this role, he became increasingly exposed to Deep Learning applications for vision applications in biology. In 2019, he started to lead a group of Helmholtz AI consultants at Helmholtz-Zentrum Dresden-Rossendorf. The team's has a mandate to help scientists from the research field matter within the Helmholtz association to use machine learning in experiment and theory. Peter is an avid open-source enthusiast, reproducibility evangelist and volunteer lecturer at TU Dresden.
Sebastian Starke: Sebastian graduated from Otto-von-Guericke University Magdeburg (Germany) as a trained mathematician and statistician. He received his PhD in biostatistics in 2024 from TU Dresden (Germany) where he was part of the 'Modeling and Biostatistics in Radiation Oncology' group at Oncoray led by Prof. Steffen Löck. During this time, he worked in the context of personalized radiotherapy, focussing on the task of improving outcome predictions of cancer patients by combining statistical survival analysis with medical image-based deep learning approaches. Since 2020, Sebastian has also been part of the Helmholtz AI consultant team led by Peter Steinbach at HZDR, where he supports Helmholtz scientists in their (mostly imaging-related) machine learning endeavours like denoising, segmentation, regression or uncertainty estimation.
Target Audience
- learners registered in a TUD masters program (bachelor students are allowed if they consider the level of the course appropriate)
- learners from any natural science domain (physics preferred, biology and chemistry encouraged)
- learners from related domains such as computer science, engineering, digital humanities, etc or any graduate studies (e.g. PhD students) of these domains are also welcome
Prerequisites
- fundamentals of python programming (basics of numpy, virtual environments and jupyter notebooks): see here for a self-study material to brush up your python
- (suggested) fundamentals of matrix-vector algebra
- (suggested) fundamentals of statistics