Recently spiking neural networks (SNNs) have received much attention because of its rich biological significance and its power in processing spatial and temporal information. However, most existing SNNs are static due to the fixed eigenvalues of spiking generation functions, which means that neurons fire spikes in a fixed frequency when receiving constant input signals. Thereby, the static SNNs are limited in scalability. In this paper, we clarify the bifurcation relationship that dynamic eigenvalues have a great influence on the neuron excitation frequency. And then we propose the Bifurcation Spiking Neural Network (BSNN) for developing a dynamic SNN. Different from traditional static SNNs, BSNN takes a bifurcation system with time-varying eigenvalues as the basic building block, thus it has more powerful flexibility and is able to handle data with complex nonlinear structures. Experiments on wide-range tasks have been conducted, including a delayed-memory XOR task, four image recognition datasets, and 25 UCR archive. The results show that the performance of BSNN is superior to the existing static SNN models.