在所有依靠Gradient Descent和Backpropagation算法来学习的Neural Network中,普遍都会存在Gradient Vanishing Problem。Backpropagation的运作过程是,根据Cost Function进行反向传播,利用Chain Rule去计算n层之前某一weight上的梯度,从而更新该weight。而事实上,在网络层次较深的情况下,我们获得的weight梯度,随着反向传播层次的深入,会呈现越来越小的状态。从而,在靠近输出端的Layers中,weight可以被很好的更新,因为可以获得不错的gradient,而在靠近输入端的Layers中,weight则更新缓慢。

举个最简单的例子,来说明该问题。如下的神经网络有四层,每层有一个node:

我们可知w是weight,b是bias,每一层的节点输入是z,输出是a,activation function是a=σ(z),我们可以得出:

当我们已知Cost Function时,我们利用Backpropagation计算weight:

可以看到,第一层的weight梯度,依赖于之后各层activation function的一阶导数之积。而对于Machine Learning中常用的Sigmoid及tanh激励函数,其derivative图像如下:

Sigmoid的derivative是[0,0.25]的,而tanh的derivative是[0,1]的。通过上式,我们看出,通过Backpropagation求梯度时,每往回传播一层,就要多乘以一项δ‘(z),也就是说,随着向回传递的深入,梯度会呈指数级的衰减,直至缩减到0,导致前层的权重无法更新。tanh要略好于sigmoid,但依然难以解决Gradient Vanishing的问题。所以Relu Function应运而生,并且在Deep Learning方面取得了巨大成功。Relu的表达式及图形如下:

其当x>0时,derivative是1,小于0时,derivative为0。该函数很好的解决了Gradient Vanishing Problem,在大多数情况下,我们构建Deep Learning时可以使用Relu作为默认的Activation Function。

Gradient Vanishing Problem in Deep Learning的更多相关文章

  1. (转)WHY DEEP LEARNING IS SUDDENLY CHANGING YOUR LIFE

    Main Menu Fortune.com       E-mail Tweet Facebook Linkedin Share icons By Roger Parloff Illustration ...

  2. Growing Pains for Deep Learning

    Growing Pains for Deep Learning Advances in theory and computer hardware have allowed neural network ...

  3. Deep Learning Libraries by Language

    Deep Learning Libraries by Language Tweet         Python Theano is a python library for defining and ...

  4. Deep learning with Python

    一.导论 1.1 人工智能.机器学习.深度学习 人工智能.机器学习 人工智能:1980年代达到高峰的是专家系统,符号AI是之前的,但不能解决模糊.复杂的问题. 机器学习是把数据.答案做输入,规则作输出 ...

  5. This instability is a fundamental problem for gradient-based learning in deep neural networks. vanishing exploding gradient problem

    The unstable gradient problem: The fundamental problem here isn't so much the vanishing gradient pro ...

  6. Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week1, Assignment(Gradient Checking)

    声明:所有内容来自coursera,作为个人学习笔记记录在这里. Gradient Checking Welcome to the final assignment for this week! In ...

  7. 课程二(Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization),第一周(Practical aspects of Deep Learning) —— 4.Programming assignments:Gradient Checking

    Gradient Checking Welcome to this week's third programming assignment! You will be implementing grad ...

  8. Deep Learning专栏--强化学习之从 Policy Gradient 到 A3C(3)

    在之前的强化学习文章里,我们讲到了经典的MDP模型来描述强化学习,其解法包括value iteration和policy iteration,这类经典解法基于已知的转移概率矩阵P,而在实际应用中,我们 ...

  9. Deep Learning in a Nutshell: History and Training

    Deep Learning in a Nutshell: History and Training This series of blog posts aims to provide an intui ...

随机推荐

  1. selenium安装及环境搭建

    说明:安装selenium前提必须是安装好了python和pip 1.安装python 在Python的官网 www.python.org 中找到最新版本的Python安装包(我的电脑是windows ...

  2. HDU 1024 Max Sum Plus Plus (递推)

    Max Sum Plus Plus Time Limit: 2000/1000 MS (Java/Others)    Memory Limit: 65536/32768 K (Java/Others ...

  3. html5实现拖拽上传

    <html><head> <meta http-equiv="Content-Type" content="text/html; chars ...

  4. NancyFx 2.0的开源框架的使用-ConstraintRouting

    新建一个空的Web项目 然后在Nuget库中安装下面两个包 Nancy Nancy.Hosting.Aspnet 然后在根目录添加三个文件夹,分别是models,Module,Views 然后往Mod ...

  5. MySQL第五天——日志

    日志 log_error(错误日志) 用于记录 MySQL 运行过程中的错误信息,如,无法加载 MySQL数据库的数据文件,或权限不正确等都会被记录在此. 默认情况下,错误日志是开启的,且无法禁止. ...

  6. openstack stein部署手册 6. nova-api

    # 建立数据库用户及权限 create database nova; grant all privileges on nova.* to nova@'localhost' identified by ...

  7. Sql 使用游标

    DECLARE data_cursor CURSOR FOR WITH T0 AS ( SELECT COUNT(f.DeptID) SubmitCount , f.DeptID FROM biz.F ...

  8. final、finally、 finalize区别

    原创转载请注明出处:https://www.cnblogs.com/agilestyle/p/11444366.html final final可以用来修饰类.方法.变量,分别有不同的意义,final ...

  9. element-ui 里面el-checkbox多选框,实现全选单选

    data里面定义了 data:[],        actionids:[],//选择的那个actionid        num1:0,//没选择的计数        num2:0,//选中的计数  ...

  10. Proto3语法翻译

    本文主要对proto3语法翻译.参考网址:https://developers.google.com/protocol-buffers/docs/proto3 defining a message t ...