Bio-Inspired Networks and Systems

In biology, parallel information processing is enabled by sparse communication over a tremendous number of connections. One challenge in the simulation of neuroscientific models is therefore the transport of spike information - ideally (faster than) in real-time. Additionally, such models are not homogeneous, so that the way in which neurons are mapped to computational units matters. To examine these aspects, we have developed therefore network simulators at different abstraction levels. Furthermore, it is not yet fully understood how information is encoded in spike sequences, i.e. which aspects of temporal relations between spikes are important. Explorations that deal with this aspect are therefore among our current activities.
Contact
Dr.-Ing. Markus Robens
Researcher
One characteristic that differs biological information processing from most current machine learning approaches is spike-based communication. While it is under debate, whether spike based communication is essential for biological learning, it has been shown that it can enable energy efficient information encoding and transfer sustained by sparse spiking activity.
Stimulated by the Advanced Computing Architectures (IVF-ACA) project, some familiarity with the models used in theoretical neuroscience has been established, while requirements on hardware resources for their simulation have been derived. An important aspect for fast simulation of these models are physical interconnect networks. In particular, their characteristics like throughput and latencies as well as the impact of neuron to node mapping are of great importance, therefore network simulators at different abstraction levels (cycle accurate and hop-based) have been developed for their examination.
As an example, network characteristics with respect to spike communication within the Cortical Microcircuit model scaled to 33% have been determined with the cycle-accurate simulator. The Cortical Microcircuit is a popular neuroscientific model implemented in NEST that has been used in some benchmarking studies. Motivated by the results of our colleagues from the IVF-ACA project, we assumed that 256 neurons can be simulated in a computational unit (CU) at an appropriate speed. This resulted in network dimensions of 11 × 10 CUs. The simulation extended over 1 second of biological real-time, while an accelerated simulation via the envisioned hardware setup was emulated by scaled time stamps. In line with the address event representation, single flit packets were applied, which were forwarded by a multicast routing scheme. For different network topologies, i.e. mesh and torus networks, that are conceptually shown at the top of the page, maximum communication latencies have been recorded as shown in the plot to the right.

Heatmaps conveniently represent the number of routed flits by a color code as shown to the right. The inhomogeneous distribution of network load is caused by different activities of the populations within the Microcircuit Model, while the different load distributions in both plots are caused by the wrap-around connections available in the torus topology.
As can be seen, such characteristics provide valuable insights for the specification of interconnect networks in the design of hardware accelerators.

Related Publications
[1] R. Kleijnen and P. Ebrahimzadeh, “Modified Communication Networks for the Simulation of Neuromorphic Systems,” Networks 2021, virtual, USA, 5-10 Jul. 2021 [Slides]
[2] R. Kleijnen, M. Robens, M. Schiek, and S. van Waasen, “A Network Simulator for the Estimation of Bandwidth Load and Latency Created by Heterogeneous Spiking Neural Networks on Neuromorphic Computing Communication Networks,” Journal of Low Power Electronics and Applications, Vol. 12, No. 2, Apr. 2022, doi: 10.3390/jlpea12020023
[3] R. Kleijnen, M. Robens, M. Schiek, and S. van Waasen, “Verification of a Neuromorphic Computing Network Simulator Using Experimental Traffic Data,” Frontiers in Neuroscience, Vol. 16, Aug. 2022, doi: 10.3389/fnins.2022.958343
[4] M. Robens, R. Kleijnen, M. Schiek, and S. van Waasen, “NoC Simulation Steered by NEST: McAERsim and a Noxim Patch,” Frontiers in Neuroscience, Vol. 18, Jun. 2024, doi: 10.3389/fnins.2024.1371103
[5] R. Kleijnen, “NeuCoNS and Stacked-Net: Facilitating the Communication for Accelerated Neuroscientific Simulations,” Ph.D. dissertation, ZEA-2, Forschungszentrum Jülich GmbH, Germany, 2024, doi: 10.34734/FZJ-2024-06126