论文部分内容阅读
提出并分析了一种全新的反馈型随机神经网络模型 ,该模型不同于常见的Boltzmann机 ,它不直接使用随机激活函数而是采用了随机型加权连接 ,神经元为简单的非线性处理单元 .揭示了该网络模型存在惟一的收敛性平稳概率分布 ,当网络中的神经元个数较多时 ,平稳概率分布逼近于Boltzmann Gibbs分布 .另外 ,还讨论了该网络模型与Markov随机场之间的关系 ,并提出了一种新型模拟退火和Boltzmann学习算法 .网络模型被成功地应用于解决难度较大的组合优化问题和人像的自动识别 ,实验结果证实了该模型具有强大的计算能力和优异的泛化性能
A new feedback-based stochastic neural network model is proposed and analyzed. This model is different from the common Boltzmann machine in that it uses random-weighted connection instead of stochastic activation function, and the neuron is a simple non-linear processing unit. Which reveals that the network model has a unique convergence of the steady probability distribution.When the number of neurons in the network is large, the stationary probability distribution approaches the Boltzmann Gibbs distribution.In addition, the relationship between the network model and the Markov random field , And presents a new simulated annealing and Boltzmann learning algorithm.The network model has been successfully applied to solve difficult combinatorial optimization problems and automatic identification of portraits.The experimental results show that the model has a strong computing power and excellent pan Performance