Seminar Machine Learning

Module Number: EI77009

Duration: 1 Semester

Ocurrence: Winter Semester and Summer Semester

Language: English

Number of ECTS: 5

Staff

Professor in charge: Reinhard Heckel

Amount of Work

Contact hours: 45

Self-study hours: 105

Total: 150

Description of Achievement and Assessment Methods

Students will give presentations as described below.

Prerequisites (recommended)

This is a seminar course targeted to advanced master students. Ideally, the students have taken previous lectures on machine learning and signal processing. Minimal requirements are a solid background on probability and statistics and linear algebra.

Intended Learning Outcomes

Upon successful completion of the module, students know the current state of the art of deep learning for imaging problems, and some of their theoretical foundations. Moreover, students will gain the ability to critically evaluate and discuss current research in that area.

Content

Deep learning methods are an ubiquitous and highly successful part of recent approaches in medical and computational imaging and vision. Until recently, most algorithms in the imaging sciences were based on static signal models derived from physics or intuition, such as wavelets or sparse representations. Today, the best performing approaches for image reconstruction and sensing problems are based on deep learning, which learn various elements of the method including i) signal representations, ii) regularizers, and iv) entire inverse functions.

For example, even with very little or no learning, deep neural networks enable superior performance for classical linear inverse problems such as denoising and compressive sensing. Motivated by those success stories, researchers are redesigning traditional imaging and sensing systems.

However, the field is mostly wide open with a range of theoretical and practical questions unanswered. In particular, deep-neural network based approaches often lack the guarantees of the traditional methods, and while typically superior can make drastic reconstruction errors, such as fantasizing a tumor in an MRI reconstruction.

In this seminar we will discuss foundational papers and notes on deep learning with a focus on imaging. Our goal is to understand how to apply those method, to understand the theoretical reason for the success of these approaches, and to find relations between deep learning and mathematically well-established techniques in imaging science and optimization.

Topics include the architectures or structure of neural networks, training of deep neural networks, deep generative models including variational autoencoders, generative adversarial networks, invertible neural networks, applications in denoising, compressive sensing, and image recovery in general, adversarial examples, robustness, and beyond.

Teaching and Learning Methods

Each week, we will discuss one or two fundamental papers or lecture notes on deep learning, with a focus on applications in imaging. Each week, two groups of Students each consisting of one or two students present. One group forms the defense team that presents the paper and emphasizes its merit. The second group forms the offense team and criticizes the same work. Both teams have to submit a written report containing a summary of the paper and their arguments. The focus is on the technical quality of the arguments and the quality of the presentation. The course is fully online, presentations and discussions will take place via zoom.

Media

The seminar will take place via zoom.