Exploring Self-Repair in a Coupled Spiking Astrocyte Neural Network.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Neuroscience

Background:

• Astrocytes play a crucial role in modulating synaptic activity at tripartite synapses.
• Endocannabinoids facilitate localized self-repair mechanisms within neural networks.
• Existing neural network models often lack inherent self-repairing capabilities.

Purpose of the Study:

• To propose a self-repairing spiking astrocyte neural network (SANN) capable of distributed self-repair.
• To introduce and validate a novel learning rule (BSTDP) for synaptic plasticity and network self-repair.
• To demonstrate the SANN's ability to maintain learned behaviors in the presence of synaptic damage.

Main Methods:

• Development of a novel BSTDP learning rule combining STDP and BCM principles.
• Implementation of a feedforward SANN architecture with astrocyte modulation and interneuron pathways.
• Simulation of a robotic obstacle avoidance task to evaluate self-repairing capabilities.

Main Results:

• The BSTDP rule effectively establishes and maintains input-output mappings by modulating plasticity window height.
• The SANN demonstrates on-the-fly self-repair driven by postsynaptic neuron activity.
• The SANN successfully maintained learned maneuvers in a robotic obstacle avoidance task with up to 80% synaptic fault density.

Conclusions:

• The proposed SANN exhibits significant distributed self-repairing capabilities at the network level.
• The BSTDP learning rule is key to enabling both learning and robust self-repair in the SANN.
• The SANN architecture, incorporating astrocytes and interneurons, offers a promising model for resilient artificial intelligence systems.