一种机器学习中防止过拟合的Dropout优化算法

A Dropout Optimization Algorithm for Preventing Overfitting in Machine Learning

针对机器学习中深度神经网络训练时常见的过拟合问题,提出了一种防止过拟合的Dropout优化算法.Dropout算法是在每批次的神经网络训练中,忽略掉一定概率的特征检测器,让某些神经元暂时停止工作,减少神经元之间的相互作用,隐式去除网络中的神经元、阻止某些特征的协同作用来缓解过拟合.算法中选择被暂时丢弃的神经元是随机概率,而优化算法在神经网络中应用伊辛模型来识别链接能量较低的神经元,并在训练和推理中把这些神经元暂时丢弃掉,算法使模型泛化性更强,有效缓解网络训练过拟合问题.

Aiming at the common over-fitting problem of deep neural network training in machine learning,a Dropout optimization algorithm to prevent over-fitting is proposed.The Dropout algorithm neglects some probabilistic feature detectors in each batch of neural network training,makes some neurons stop working temporarily,reduces the interaction between neurons,implicitly removes the neurons in the network,and prevents the synergy of some features to alleviate over-fitting.The selection of temporarily discarded neurons in the algorithm is random probability,while the optimization algorithm applies Ising model to identify neurons with lower link energy in the neural network,and discards these neurons temporarily in training and reasoning.The algorithm makes the model more generalized and effectively alleviates the problem of over-fitting in network training.