方法有三: L1和L2正则 Dropout early stop 1. L1和L2正则: 模型权重越小,复杂度越低。越能防止过拟合 L1正则和L2正则 公式、以及反向梯度传播: https://jackaudrey.gitbooks.io/machina-learning/content/fu-ff1a-ji-qi-xue-xi-bu-chong-zhi-shi-li-jie-shuo-ming/02l1he-l2-regularization.html