Open Topics in the Area of Artificial Intelligence/Machine Learning (AI/ML)

This page lists all open topics in the area of AI/ML for communication networks. Interested students can contact the supervisor(s) to ask questions or discuss the details of a topic. It is also possible to propose your own topic in the area. An overview of AI/ML applications in networking can be found in [1].

Supervisors of our AI/ML work group: Christian Vielhaus, Johannes Busch, Vincent Latzko, Jonas Schulz

Uncertainty Quantification in Deep Learning for Transition-Aware Human Activity Recognition

Supervisor: Jonas Schulz

Human Activity Recognition (HAR) plays a crucial role in pervasive and ubiquitous computing, tactile internet as well as healthcare and entertainment. By leveraging the knowledge of human activities, network resources can be pre-allocated and services provided appropriately.
Inertial sensors are often used to acquire information about human movement patterns when these sensors are attached to the human body. Existing approaches make use of advancements in deep learning to detect and recognize human activities. Still, they often do not consider the deployment of these classifiers in the wild which comes with multiple challenges. One is, that humans are not restricted to performing activities (such as walking, running or sitting) pre-defined in a dataset. To recognise cases, in which a human is performing arbitrary activities not defined in a dataset is still an open challenge. One way to tackle this particular problem is to monitor the uncertainty of deep learning-based classifiers and to exploit this additional information in combination with thresholding to detect so-called out-of-distribution (OoD) samples. To get a truthful estimate of the classifier’s uncertainty, deep ensembles can be employed.
Motivated by this challenge, the framework of this work can be formulated as uncertainty quantification in deep learning for transition-aware human activity recognition based on ensembling methods. The thesis will involve implementing state-off-the-art deep ensembles and analysing how incorporating knowledge about the model uncertainty can be used to detect OoD samples on benchmark datasets.

  • Starting time: Immediate
  • Student and Diploma/Master thesis (with task extension)
  • Required skills: basic knowledge of machine/deep learning and signal processing
  • Desired skills: Python, Deep Learning Frameworks such as Tensorflow, Pytorch etc ..

Motivation Links:

Joined Optimization of Traffic Signaling and Vehicle Speed Advisory in V2I-Enabled Traffic

Supervisor: Johannes Busch

With emerging Vehicle-to-Infrastructure (V2I) communication technologies, such as IEEE 802.11p or 5G, arises the opportunity for an intelligent traffic infrastructure that is able to quickly adapt to the current state of the traffic system.Through the exchange of information between individual vehicles and the traffic infrastructure, traffic signaling could be tailored to individual vehicles. Furthermore the traffic infrastructure could advise drivers on appropriate driving velocities. These measures could improve traffic flow and therefore mitigate environmental and economic repercussions of traffic systems. However, the optimization of large distributed systems under the availability of intricate knowledge about the system state is not trivial and novel control paradigms need to be explored. A promising candidate could be Deep Reinforcement Learning (DRL), which has shown to be able to solve many intricate control problems by learning from interaction and has also successfully been applied to traffic control systems.

In this Diploma/Master thesis, a joint optimization of traffic signaling and speed advisory with DRL should be explored. Subtasks include:

  • A literature review of the current state of the art of intelligent traffic control, V2I, and DRL.
  • An extension of an existing traffic simulator to enable the joint control of traffic signaling and speed advisory, under the knowledge of the current state of the traffic system.
  • The optimization of the developed simulation with an appropriate DRL algorithm. Preferably, the proposed solution should be scalable to large infrastructures.
  • An experimental review of the benefits of the developed intelligent V2I-enabled traffic control system. Various experimental setups could be explored here. For instance, the scalability to large traffic systems or the influence of different penetration rates of V2I-enabled vehicles could be two particularly interesting paths to explore.

Prospective students should have previous knowledge in the Python programming language and a keen interest and previous experience in Machine Learning (preferably DRL).

  • Starting time: Immediately
  • Diploma/Master thesis
  • Required skills: Python, experience with ML

Motion Capturing & Data Analysis in Physiotherapy

Supervisor: Giang Nguyen

The vision of the spin-off project Veiio in Dresden is to develop an intelligent suit with fully integrated sensors and actuators that uses motion capture to provide real-time vibrotactile feedback to the wearer.

We offer a diploma thesis project in Motion Capture (MoCap) Data Analysis based on Inertial measurement unit (IMU) sensors. In this project, you should develop a prototypical system that recognizes selected movement exercises and that evaluates the movement execution in real-time. Included tasks are also: acquiring the data with provided hardware, testing, and comparing different approaches for motion classification and evaluation, providing a simple interface for examining the algorithms.

You should bring:

  • Profound knowledge in Python programming
  • Experience with approaches from machine learning (e.g., scikit-learn, PyTorch)

We provide:

  • Application-oriented project that is intended to be used
  • Continuous exchange with interdisciplinary team

That sounds exciting?
Then please contact us: info@veiio.de or Giang Nguyen

References

[1] A Comprehensive Survey on Machine Learning for Networking: Evolution, Applications and Research Opportunities
[2] Network congestion