Towards Biology-Inspired Fault Tolerance of Neuromorphic Hardware for Space Applications

High-energy particles in space can induce single-and multibit upsets in random electronic components of FPGA-based neuromorphic systems. We propose NeuFT, a low overhead biology-inspired architecture for fault tolerance of these systems. NeuFT draws inspiration from astrocytes, which are star-shaped glial cells in a mammalian brain that facilitate the self-repair of failed neurons by sending a closed-loop retrograde feedback signal. Our fault-tolerant design methodology is as follows. First, we partition a spiking neural network (SNN) model into synaptic islands and place an astrocyte in each of them. Each astrocyte is trained using spike trains of neurons on its synaptic island to optimize fault tolerance capabilities. Next, we propose a practical astrocyte-inspired controller that integrates astrocyte feedback as bias of a leaky integrate and fire neuron with minimal design changes. Finally, we introduce software to partition and train an astrocyte-enabled SNN and map it to hardware. We evaluate NeuFT for five datasets with varying degrees of bit error rates considering single- and multibit upsets. Our results show a significant reduction in fault tolerance overhead while delivering a high classification accuracy with respect to state-of-the-art.