Use Cases
Explore the dynamic and interdisciplinary research groups shaping progress across a wide range of fields
Enabling Synthetic Biological Intelligence for Future AI Computation Biological computing
The work in this research area focuses on utilizing simple toy examples to showcase the feasibility of mapping computing tasks solvable on digital hardware to alternative hardware, such as biological hardware. Therefore, an overview of the start-of-the-art will be provided, outlining possible research gaps. For an evaluation of the benchmarked biological computing platform–a cluster of real neurons on a silicon chip– the work focuses on pre-defined key performance indicators, such as latency and sample efficiency.
Further investigation points of the biological hardware are the interface, the interaction with the system, and the integration in existing computing infrastructure.
Overall, the approach is not limited to practical approaches but also includes theoretical approaches, such as mathematical models for signal transmission.”
Members:
Yannik Böck
Adalbert Fono
Pit Hofmann
Analog computing
Our research group investigate energy-efficient neural computation using both digital and analog in-memory computing hardware. To enhance energy efficiency and accuracy, we study sparsified models on digital platforms, such as Spinnaker and AxeleraAI. In parallel, we explore analog computing to assess different neuron models for spiking neural networks and their impact on energy-accuracy tradeoffs. We will utilize the available hardware from Anabrid, Axelera, Spinncloud, and other researchers to simulate and control cyber-physical systems, evaluating key metrics such as latency, accuracy, and energy consumption. Our methods involve deploying AI models on hardware compatible with PyTorch and comparing sparsified models from LMU against conventional ones. A final demonstrator will benchmark these models across GPUs, neuromorphic systems, and emerging architectures.
Members:
Dr. Juan A. Cabrera
Manjot Singh
Fatima Rani
Sifat Rezwan
Volkan Ün
Isikcan Yilmaz
Energy efficiency of neuromorphic Hardware
Our research area explores the potential for efficiently deploying Spiking Neural Networks on neuromorphic hardware, with a focus on medical applications. We examine key design choices and computational strategies that could enhance performance and energy efficiency in these systems. Our goal is to better understand how these biologically inspired models might offer responsive and resource-conscious solutions for healthcare technologies.
Members:
Dr. Ernesto Araya Valdivia
Claas de Boer
Duc Anh Nguyen
Miriam Kranzlmüller
Alexandru-Paul Drăguţoiu
Functional compression
Our research group centers on leveraging functional/semantic compression techniques for 3D medical image reconstruction over networked environments, with a focus on implementing these methods across both analog and digital hardware platforms. We plan to replace voxelization-based color-coding algorithms with innovative deep learning approaches, such as the information bottleneck principle, to optimize functional/semantic compression.
Additionally, we aim to investigate biologically inspired computational models to advance medical imaging and communication systems.
Members:
Chengyang Li
Dr. Jianfei Li
Sifat Rezwan
Diffusion Models on neuromorphic Hardware
Our research goal is to enable real-time, low-power control for robotic systems. In this group, we explore the use of diffusion policy models for tasks in surgical robotics. The core innovation of our work is to combine compression techniques such as quantization and pruning, with lightweight architectures like Mixture of Experts, to run advanced imitation learning policies on the efficient SpiNNaker2 neuromorphic hardware.
Members:
Dr. Massimiliano Datres
Yin Li
Lorenzo Mazza
Sarah Pardo
Jonas von Berg
Martin Lelis