Increasing Simulated Radiation Robustness of Graph Spiking Neural Networks Leveraging brain-adaptation mechanisms in a Minimum Dominating Set Approximation Algorithm

Toeter, Akke

Increasing Simulated Radiation Robustness of Graph Spiking Neural Networks Leveraging brain-adaptation mechanisms in a Minimum Dominating Set Approximation Algorithm

Files

Toeter, A. s-1047775-MSc-MKI92-Thesis-2023.pdf (7.06 MB)

Authors

Toeter, Akke

Issue Date

2023-07-16

Language

en

Abstract

Neuromorphic architectures are of interest in space application due to their energy efficiency. However, space radiation has been shown to damage computational hardware. Research has been performed on the radiation robustness of (pre-trained) spiking neural networks, and the brain has shown to be able to recover from certain types of lesions. Other work has shown Spiking Neural Networks (SNNs) can obtain advantages for graph algorithms. Such (distributed) SNNs graph algorithms may become relevant for space applications. For example, SNNs may form a neuromorphic alternative to the Minimum Dominating Set (MDS) algorithm that others have proposed for distributed satellite swarm coordination. This research combines the trend of SNN implementations of distributed graph algorithms with neuromorphic space applications and brain-adaptation induced robustness as inspiration. It shows that brain inspired adaptation mechanisms can increase the simulated radiation robustness of an SNN implementation of an unweighted, distributed minimum dominating set approximation (MDSA) algorithm. An SNN implementation of the MDSA algorithm by Alipour et al. is created and used for this experiment. That SNN is adapted with a population coding approach, and with a sparse redundancy approach. These three SNNs are exposed to simulated radiation effects in the form of synaptic weight increases which occur with a probability of 0.001% to 20% per synapse per timestep. Separate simulations are performed with a simulated radiation induced permanent neuron death with a probability of 0.01 % to 25 % per neuron per timestep. The SNN performance is measured by comparing its output to the unradiated algorithm output. The sparse redundancy increases the robustness against simulated radiation induced neuron death for radiation probabilities from 0.5 % to 25% per neuron per time step, for the MDSA SNN. Below these probabilities, the adaptation mechanism is contra productive. Similarly, population coding is contra productive below 0.1% and increases radiation robustness for simulated synaptic weight increase probabilities of up to 5%.