fortiss is the research institute of the Free State of Bavaria for software-intensive systems and services with headquarters in Munich. The institute currently employs around 150 employees, who collaborate on research, development and transfer projects with universities and technology companies in Bavaria, Germany and Europe. Research is focused on state of the art methods, techniques and tools of software development, systems & service engineering and their application to reliable, secure cyber-physical systems, such as the Internet of Things (IoT). fortiss has the legal structure of a non-profit limited liability company (GmbH). Its shareholders are the Free State of Bavaria (as majority shareholder) and the Fraunhofer Society for the Promotion of Applied Research.
This master’s thesis will be jointly supervised by the Neuromorphic Computing team at fortiss and the ACES Lab at the Chair of Theoretical Information Technology.
In the design of wireless communication systems, models, such as channel models, have traditionally played a crucial role. However, these models always represent an idealized simplification and cannot fully capture all practical complexities. In contrast, data-driven approaches based on deep neural networks eliminate the need for explicit modeling and therefore have the potential to outperform classical designs. Current research explores deep learning techniques for tasks such as channel estimation, signal detection and modulation / demodulation, channel encoding and decoding, and physical-layer authentication.
Spiking neural networks (SNNs) are a novel type of neural network and differ fundamentally from conventional deep learning models by incorporating temporal dynamics and event-driven computation. SNNs operate using discrete spikes, similar to biological neurons, offering advantages in energy efficiency and temporal processing. Due to their event-driven, energy-efficient nature, SNNs are gaining attention in communications research.
This thesis focuses on designing and implementing SNNs to address selected physical-layer tasks, such as channel estimation. Initially, the performance of the implemented SNN will be evaluated through simulations of the SNN on GPUs. Subsequently, the network will be deployed on the best suitable specialized hardware platform (such as SpiNNaker2, Pulsar, Loihi 2, Akida 2) to assess real-time performance under practical constraints, including latency and energy consumption. Furthermore, data from USRP X410 software-defined radios (SDRs) will be utilized to study the performance on real-world data.
