부산대학교 도서관

Information processing by spiking neural networks (SNNs) is one of the greatest applications of neuroscience research. The benefits of biologically inspired SNNs are known for energy efficient computations through spike-driven communications. However, the biological relevance of existing computational models and hardware implementations is rather limited. It is known that synaptic transmission in a living brain is directly influenced by astrocytes releasing gliotransmitters that modulate the excitability of neurons and, hence, their firing rate. Unlike electrical spikes with a shape determined by the properties of a neuron, the amplitude of the astrocyte’s response is gradual (proportional to the input stimulus). In the presented study, we use this feature for non-binary information processing. We employ SNN enhanced by bidirectional interaction with an astrocytic network to recognize grayscale images, encoded into astrocyte activation levels. The results showed that such a harmony of digital and analog coding makes it possible to retrieve even highly noisy images within a few seconds. This memory effect is provided only by astrocytes, and the storage time is determined by the characteristic time scale of their activation.