原ppt下载:pan.baidu.com/s/1nv54p9R,密码:3mty

 

需深入实践并理解的重要概念:

Deep Learning:

  • SoftMax Fuction(输出层归一化函数,与sigmoid相似的激活函数,用于解决分类问题(分类大于2时;sigmoid解决二分类问题))

          1)读《Deep Learning Tutorial》(台湾大学 李宏毅 深度学习教学ppt)后杂记

          2)每个neuron的softmax输出:读《Deep Learning Tutorial》(台湾大学 李宏毅 深度学习教学ppt)后杂记,其中:读《Deep Learning Tutorial》(台湾大学 李宏毅 深度学习教学ppt)后杂记

DNN(Deep Neural Networks):

  • MSE(Means Square Error,均方误差) / CE(Cross Entropy,交叉熵)

          - Use to minimum total costs for softmax layer. CE is better.

          - MSE minimum:读《Deep Learning Tutorial》(台湾大学 李宏毅 深度学习教学ppt)后杂记

          - CE minimum:读《Deep Learning Tutorial》(台湾大学 李宏毅 深度学习教学ppt)后杂记

  • Mini-batch & batch_size(decides how many examples in a mini-batch) & epoch(周期)

          - batch:样本训练中,将完整数据分为等量的多个batch(批次),每次输入一个batch而不是完整样本进行训练

          - epoch:周期被定义为向前和向后传播中所有batch的单次训练迭代

          - mini-batch has better performance than original gradient descent

  • Vanishing Gradient Problem(梯度消失问题)
  • ReLU(Rectified Linear Unit,线性纠正单元)

          - As an activative function, used when the number of layers is quite large.

          - 对于大于0的所有输入来说,它都有一个不变的导数值;常数导数值有助于网络训练进行得更快,常用于多层隐藏层

          - Special cases of MaxOut:读《Deep Learning Tutorial》(台湾大学 李宏毅 深度学习教学ppt)后杂记

  • Learnable activation function
  • Adaptive learning rate(学习率:每次迭代中cost function中最小化的量。简单来说,我们下降到cost function的最小值的速率是学习率

          - Use a large rate first, then change to a small one

  • Momentum(动量原理)

          - Use the optimizer Adam(Advanced Adagrad Momentum)

  • Overfitting Problem(过拟合问题)

          - Use early stopping

  • Weight Decay(训练时用p%的dropout,测试时对权值做(1-p%)的调整后再获得输出)
  • Dropout(训练的过程舍弃神经元)

          - Will change structure of networks while training. better than MaxOut

CNN(Convolutional Neural Networks):

  • Image recognization suits to use CNN because of 3 important properties

          1) Patterns are much smaller than the whole image

          2) The same patterns appear in different regions

          3) Subsampling pixels does not change the object

  • filter & channel
  • stride(step)
  • zero-padding
  • max-pooling
  • flattern
  • less parameters because of sharing weights