论文部分内容阅读
针对DBN网络隐含层层数难以选择的问题,首先从数学生物学角度分析了随机初始化的梯度下降法导致网络训练失败的原因,并进行验证,证明了RBM重构误差与网络能量的正相关定理;然后根据隐含层和误差的关系,提出一种基于重构误差的网络深度判断方法,在训练过程中自组织地训练网络,使其能够以一种接近人类处理问题的方式解决AI问题.手写数字识别的实验表明,该方法能够有效提高运算效率,降低运算成本.
In order to solve the problem that the layer number of hidden layers in DBN is hard to choose, this paper first analyzes the reason that the gradient descent method of stochastic initialization leads to the failure of network training from the perspective of mathematical biology, and verifies that the RBM reconstruction error is positively correlated with network energy According to the relationship between hidden layer and error, a method of network depth judgment based on reconstruction error is proposed to train the network in a self-organizing manner so that it can solve the AI problem in a way that is close to the human processing problem Experiments on handwritten numeral recognition show that this method can effectively improve the computational efficiency and reduce the computational cost.