Effect Exploration of the Spiking Neuron Property and Heterogeneity on Brain-Inspired Context-Dependent Learning

Brain-inspired computing uses the information processing method of human brain to process unstructured information in real time. It is an important way to artificial general intelligence. Spiking neural network (SNN), as the most representative network framework for brain-inspired computing, can realize advanced cognitive functions. Context-dependent learning is to retain the cognition of old knowledge when learning new knowledge, make flexible response to changing situational information in the environment, so as to achieve advanced cognitive functions. This study explores how neuron property and network structure heterogeneity affect SNN with winner-take-all mechanism. It directly determines the performance of context-dependent learning. This study shows that the property of neurons and the heterogeneity of network structure play a decisive role in the effectiveness of context-dependent learning. Results reveal the membrane capacitance and conductance affect the charging and discharging time of neuron cells and thus change the learning convergence rate. We summarize the effect of distance between the threshold of membrane potential and the resting potential on the overall accuracy of learning. Moreover, the addition of Gaussian noise to simulate the small fluctuations in the SNN shows that the heterogeneous network can improve the learning ability of neurons. This work can be potentially used in intelligent navigation, low-power edge devices, unmanned system, and neuro-robotics.