Development of a Bio-Inspired Computational Astrocyte Model for Spiking Neural Networks

The mammalian brain is the most capable and complex computing entity known today. For many years there has been research focused on reproducing the brain\u27s processing capabilities. An early example of this endeavor was the perceptron which has become the core building block of neural network models in the deep learning era. Deep learning has had tremendous success in well-defined tasks like object detection, games like go and chess, and automatic speech recognition. In fact, some deep learning models can match and even outperform humans in specific situations. However, in general, they require much more training, have higher power consumption, are more susceptible to noise and adversarial perturbations, and have very different behavior than their biological counterparts. In contrast, spiking neural network models take a step closer to biology, and in some cases behave identically to measurements of real neurons. Though there has been advancement, spiking neural networks are far from reaching their full potential, in part because the full picture of their biological underpinnings is unclear. This work attempts to reduce that gap further by exploring a bio-inspired configuration of spiking neurons coupled with a computational astrocyte model. Astrocytes, initially thought to be passive support cells in the brain are now known to actively participate in neural processing. They are believed to be critical for some processes, such as neural synchronization, self-repair, and learning. The developed astrocyte model is geared towards synaptic plasticity and is shown to improve upon existing local learning rules, as well as create a generalized approach to local spike-timing-dependent plasticity. Beyond generalizing existing learning approaches, the astrocyte is able to leverage temporal and spatial integration to improve convergence, and tolerance to noise. The astrocyte model is expanded to influence multiple synapses and configured for a specific learning task. A single astrocyte paired with a single leaky integrate and fire neuron is shown to converge on a solution in 2, 3, and 4 synapse configurations. Beyond the more concrete improvements in plasticity, this work provides a foundation for exploring supervisory astrocyte-like elements in spiking neural networks, and a framework to implement and extend many three-factor learning rules. Overall, this work brings the field a bit closer to leveraging some of the distinct advantages of biological neural networks