Spike Neural Network with Delayed Propagation Characteristics and Hardware Implementation
Gaoyuan Wang, Dongmei Fu · 2024
Spike neural networks, as mathematical models for simulating biological neurons, have received high attention in the fields of life sciences and artificial intelligence due to their biological rationality and hardware friendliness in working mechanisms. However, the discrete signal form of pulse neurons makes them non differentiable and unable to propagate backwards. This paper proposes a delayed propagation pulse neuron model to solve this issue. This model operates on the long-term characteristics of pulse neurons, by treating the effect of pulse signals on neurons as a continuous charging current to achieve signal continuity and thus obtain differentiable characteristics, achieving backpropagation. At the same time, the use of capacitance charging models to construct low-cost neuron hardware circuits provides a new development direction for brain-like computing hardware.