Exploring Self-Repair in a Coupled Spiking Astrocyte Neural Network

Abstract

It is now known that astrocytes modulate the activity at the tripartite synapses where indirect signaling via the retrograde messengers, endocannabinoids, leads to a localized self-repairing capability. In this paper, a self-repairing spiking astrocyte neural network (SANN) is proposed to demonstrate a distributed self-repairing capability at the network level. The SANN uses a novel learning rule that combines the spike-timing-dependent plasticity (STDP) and Bienenstock, Cooper, and Munro (BCM) learning rules (hereafter referred to as the BSTDP rule). In this learning rule, the synaptic weight potentiation is not only driven by the temporal difference between the presynaptic and postsynaptic neuron firing times but also by the postsynaptic neuron activity. We will show in this paper that the BSTDP modulates the height of the plasticity window to establish an input-output mapping (in the learning phase) and also maintains this mapping (via self-repair) if synaptic pathways become dysfunctional. It is the functional dependence of postsynaptic neuron firing activity on the height of the plasticity window that underpins how the proposed SANN self-repairs on the fly. The SANN also uses the coupling between the tripartite synapses and ɣ-GABAergic interneurons. This interaction gives rise to a presynaptic neuron frequency filtering capability that serves to route information, represented as spike trains, to different neurons in the subsequent layers of the SANN. The proposed SANN follows a feedforward architecture with multiple interneuron pathways and astrocytes modulate synaptic activity at the hidden and output neuronal layers. The self-repairing capability will be demonstrated in a robotic obstacle avoidance application, and the simulation results will show that the SANN can maintain learned maneuvers at synaptic fault densities of up to 80% regardless of the fault locations.

Autonomous Learning Paradigm for Spiking Neural Networks

“…However, after a period of learning, only one actuator neuron will be active (left turn motion) due to the potentiated synaptic weight, and the other actuator neuron becomes inactive as the associated synaptic weight is depressed. Compared to other approaches such as [7], the proposed method can learn and adapt to the surrounding environmental conditions based on the STDP kernel structures. Therefore, the proposed learning approach does not require an input to output mapping table and thus points to a possible future direction for SNN metaplasiticity.…”

“…However, the requirement for a supervised approach constrains the design, development and deployment of the SNN systems, especially for applications operating within a dynamic environment, e.g. robots [7].…”

“…A combined STDP/BCM learning rule, termed BSTDP, has be demonstrated in an SNN-based robotic controller application to implement learning, e.g. in the approach of [7] the spiking astrocyte neural network used BSTDP to implement both learning and self-repair in the robotic applications. However, it required supervisory signals to achieve the correct input/output mapping.…”

Hardware-Intrinsic Multi-Layer Security: A New Frontier for 5G Enabled IIoT

“…The output layer neuron provides a stable enable signal for the client FPGA if the received device parameters are within scope. This principle of using a spiking neural network is derived from [32,33], and hardware realization of the approach is described in [34].…”

“…However, in [32][33][34], the authors derive bio-inspired principles for homeostasis targeting robotic applications, where this paper emphasis the use of similar methodologies for hardware Trojan detection. Bio-inspired computing develops computational models using various models of biology.…”