训练深度神经网络失败的罪魁祸首不是梯度消失,而是退化 https://mp.weixin.qq.com/s?__biz=MzA3MzI4MjgzMw==&mid=2650735824&idx=1&sn=1720a687f197538b23d8d329e87998fb&chksm=871ac0aeb06d49b826d8f7d6370e2fc9768c57c4d8ef3ed8dd90ae97d33bbf5fb50d0c6a6d03&mpshare=1&scene=23&srcid=0107R95kMPAtSkgppqJnBLOb#rd