Forschungsthemen

• Analog Tunability of Memristive Devices: A Physics- and Machine Learning-Driven Approach
  The steep non-linearity of memristive switching dynamics, coupled with inherent variations and asymmetry, presents significant challenges in achieving accurate and efficient control of analog conductance updates. By integrating electrical characterization, device physics, and machine learning techniques, we develop tailored models to analyze switching dynamics and optimize schemes for precise multi-level tunability. Effective implementation of analog tunability in memristive devices constitutes a crucial enabler for analog and adaptive in-memory computing. (Collaboration with PGI-7 and PGI-15)
• Multi-Timescale Neuromorphic Computing & ML
  This research is inspired by the biophysics of locally-activated and long-lasting dendritic spikes, triggered by synaptic inputs. This mechanism integrates two distinct timescales: rapid stimulus detection through synaptic transmission and slower integration of these events via long-lasting memory traces in the local membrane potential. We explore computing task improvements through hardware co-design inspired by this biological mechanism (Collaboration with Institute of Cognitive Science, Osnabrück). As a benchmark for this brain-inspired approach, we investigate the hardware implementation of specific machine learning algorithms, such as the Structured State-Space Sequence (S4/SSM) models, known for its effectiveness in capturing long-range dependencies in streaming data
  (Collaboration with Dr. Sebastian Siegel and PGI-15).
• Multi-Task Neuromorphic Processing & ML
  Beyond context-independent feedforward propagation with synaptic weights, dendrites play a crucial role in facilitating contextual neuronal modulation within hierarchical sensory processing pathways, enabling multi-task representation learning. We investigate the computational benefits of hardware co-design inspired by this mechanism (Collaboration with Dendritic Learning Group, PGI-15). As a benchmark for this brain-inspired approach, we explore the hardware implementation of specific machine learning algorithms such as Model-Agnostic Meta-Learning (MAML) and Memory-Augmented Neural Networks (MANN), known for their promise in few-shot and continual learning scenarios.