Abstract: Although the artificial neural network is almost comparable to the human brain in image recognition,the activation functions such as ReLU and Softplus are only highly simplified and simulated for the output response characte-ristics of biological neurons.There is still a huge gap between the artificial neural network and the human brain in many aspects,such as noise resistance,uncertainty information processing and power consumption.In this paper,based on the simulation experiments of biological neurons and their response characteristics,a strong anti-noise activation function Rand Softplus with biological authenticity was constructed by defining and calculating parameters η,which reflects the randomness of each neuron.Finally,the activation function was applied to the depth residuals network and verified by facial expression dataset.The results show that the recognition accuracy of the activation function proposed in this paper is almost equal to the current mainstream activation function when there is no noise or a small amount of noise,and when the input contains a large amount of noise,it shows good anti-noise performance.

Key words: Activation function, Anti-noise, Leaky integrate-and-fire model, Neural networks