Self-repairing Memristive Spiking Astrocyte Neural Network HardwareArchitecture

Recent research indicates that bi-directional coupling between neurons and astrocytes gives rise to a self-repairing capability in the human brain. In this paper based on the model of emerging nanodevice memristor, a novel self-repairing spiking astrocyte neural network (SANN) is proposed which can mimic the brain neural net to demonstrate a fine-grained self-repair capability for the hardware circuits. A learning phase state control circuit (hereafter referred to as the LPSC circuit) is embedded in this memristor-based SANN (MSANN). Through the LPSC circuit, an input-output mapping can be established during the learning phase, and this mapping is maintained via self-repair process if synaptic pathways become damaged or broken. We will show that the synaptic weights are modulated by the firing activities of pre- and postsynaptic neurons through the LPSC circuit, which is consistent with the Bienenstock, Cooper, and Munro learning rule. The interaction between the tripartite synapses and γ-GABAergic interneurons is captured in the MSANN circuit that gives rise to a presynaptic neuron frequency filtering capability on the synaptic sites. This provides the network an ability to route information (represented as spike trains) to different neurons in the next layer of the MSANN. The self-repairing performance of the network is evaluated by a typical robotic application of obstacle avoidance, where different synaptic fault densities are considered. Results show that the functionality of the proposed MSANN can be maintained even under an extreme failure condition of >80%. This provides an alternative robust neuromorphic hardware system design for critical task applications.