After successful completion of the course, students are able to:

• Understanding the principles of Deep Learning and recognize suitable problems which are solvable with Deep Learning

• Estimate and execute the organisational tasks involved in data science projects.

• Solving a specific research task with Deep Learning (e.g., object detetion, intelligent agents, or natural language processing).

• Selecting a suitable Deep Learning model for the problem at hand and training it efficiently

• Assess the found solution and present the results appropriately.

The focus of this course is a project, which has to be solved applying Deep Learning methods. The topic of that project can be chosen by the student. To be able to complete the project within this course, weekly lectures cover the most essential methods. Apart from an overview over Deep Learning and Neural Networks, the following advanced topics will be presented in this course and the project must use at least one of them:

• Convolutional Neural Networks for Image Analysis

• Recurrent Neural Networks for Sequence modeling

• Deep Reinforcement Learning

• Autoencoders and Deep Generative models

• Transformers

• Graph Neural Networks

• Explainable AI

Additionally, practical aspects are presented throughout the course, like software libraries and frameworks, which aid the implementation process and allow to communicate the results from the project by presenting them in a meaningful way.

The weekly lectures will cover the theory and applied deep learning methods, as well as practical tips to successfully realize the student project.

The project is divided into three phases that are graded separately:

1. Selection and formulation of a suitable problem as well as procuring a suitable dataset. The goal is to find and investigate an interesting and challenging problem, for which other approaches might not work as well as Deep Learning. Students are free to choose a problem from different areas, e.g., computer vision, intelligent agents, machine translation, or audio processing. Once the problem has been formulated, a suitable dataset needs to be assembled. Depending on the question under investigation, an existing dataset can be re-used or a new dataset has to be collected.

2. Selecting and applying a suitable model to process the dataset. In this phase, students are supposed to implement their solution. They have to select appropriate tools to efficiently train and optimize a complex model.

3. Assessment and presentation of the solution. To assess the found solution, it has to be compared to scientific work that represents the state of the art. Finally, the project should be prepared in such a way that potential users could use it, e.g., via an API or a simple mobile application.

At the end of the project, a final report has to be compiled which contains the results from those four phases. A short overview is also presented in class at the end of the semester.

Exercises have to be solved by each student individually.

ECTS Breakdown: 3 ECTS = 75h
16h Lecture
45h Programming exercise
10h Creating the final report and the presentation
  4h Present the final results
-------------------------------------------------------------------------------
75h Total workload

The proof of accomplishment consists of three parts. A software development project that investigates and attempts to solve a particular problem with Deep Learning, a technical report, and the presentation of the results.

The project is divided into three parts that are graded separately. Students are graded on their understanding of the underlying theory and their competence in solving a given problem independently.

The results are presented in the last two lectures.

The final registration takes place during the first lecture: Places will be allocated among those present in the order of the initial registration in TISS.

Deep Learning - Goodfellow et al.

Taking the course Deep Learning for Visual Computing before this course is highly recommended, as it covers the fundamentals of Deep Learning (neural networks, optimization, backpropagation, etc.) in much greater depth, and applies them on a specific problem from the domain of Visual Computing.