前向神经网络容错性分析的切比雪夫不等式法

Fault-tolerance analysis of feedforward neural networks with the Chebyshev inequality method

目前 ANN的分析中缺乏对硬故障容错性能的分析 ,针对这一问题利用切比雪夫不等式给出了一种容错性分析的估算方法。利用切比雪夫不等式 ,分析了具有可微作用函数的前向神经网络容错性 ,建立了前向神经网络随机故障模型 ,讨论了固定型连接故障和错误输入故障对单个神经元的影响 ,通过分析这种前向神经网络故障传播特点 ,结合神经元容错分析的结论 ,得出了前向神经网络容错性分析的算法和相应公式。通过仿真实验 ,验证了上述结论的正确性。

Current analyses of ANN seldom consider difficult faults. This paper presents a fault tolerance evaluation method for difficult faults based on the Chebyshev inequality. The Chebyshev inequality is used to analyze the fault tolerance of feedforward neural networks with differentiable activation functions. A stochastic fault model for feedforward neural networks was then built. The effects of stuck at faults and error inputs on the neurons were also discussed. An algorithm and a corresponding formula for fault tolerance analysis of feedforward neural networks are presented based on the features of fault propagation. Computer simulations verify the theoretical analysis.