Effects Investigation of Critical Spiking Neuronal Feature on Dendritic Credit Assignment for Efficient Spike-based Learning

As a model inspired by biological neural network mechanism, spiking neural network plays an important role in realizing brain-inspired computing with higher cognitive functions. The credit assignment problem is the basis of the internal decision-making process of network learning. Its solution directly determines the learning performance of spiking neural network. This paper focuses on the influence of features of spiking neurons on credit assignment processes with dendritic compartments. We study the calculation and transmission of feed-forward and feedback signals in dendritic credit assignment. The results show that the changes of membrane potential, leakage conductance and dendrite conductance affect the credit assignment process. Credit assignment is closely related to the error calculation of feedforward signals and feedback signals in network learning, which makes its change affect the learning performance. This is due to these features acting on the plateau potential and weight updating. The results also reveal the effects of long time constant, short time constant and maximum neuron firing rate on credit assignment. These features affect learning performance by the response kernel of the error calculation process that acts on credit assignment. The exploration of these critical features in this work is helpful for improving network learning performance. This work can be applied in areas such as smart healthcare, intelligent navigation and edge computing.