Publications

Translating novel insights from computational neuroscience into neuromorphic computing, we show how tightly balanced Spiking Neural Networks outperform existing networks in efficiency, robustness while maintaining competitive accuracy.

NoisyDECOLLE is a Python framework for assessing the robustness of SNNs trained with local learning rules inspired by three-factor learning and synaptic plasticity.

Designing and implementing the neuroAIx FPGA cluster, we show how a combination of novel communication topology, local synchronization algorithm and design paradigms like maximizing neuron-per-node density lead to a neuromorphic system that enables high-speed, deterministic and efficient neuroscience simulations.

The neuroAIx framework is a collection of C++-based simulation tools and hardware platforms to evaluate novel hardware concepts for neuroscience experiment accelerators.

We derive an approach on how to systematically design cost-efficient BNNs, with novel methods like hybrid ternarization and a hardware-based cost estimation leading to BNNs more efficient that existing ones without compromising accuracy.