10 - Convolutional Networks
这节课包括三部分：
- 卷积网络的引入。从全连接层在处理图像任务上的缺陷引入卷积，然后介绍卷积的优势，以及卷积的计算细节。
- 卷积在实践中常规操作：
- Padding。为了使得输入输出图像大小一致
- Strided Convolutions / Pooling。解决难以捕捉不同尺度下的视觉信息的问题
- Dilations。增大感受野
- 群组卷积。降低卷积操作的计算复杂度
- 卷积的微分
Convolutional operators in deep networksThe problem with fully connected networksHow convolutions “simplify” deep networksAdvantages of convolutionsConvolutions in deep networksElements of practical convolutions1. Padding2. Strided Convolutions / PoolingGrouped ConvolutionsDilationsDifferentiating convolutionsWhat is needed to differentiate convolution?Refresher on differentiating matrix multiplicationConvolutions as matrix multiplication: Version 1The adjoint of a convolutionConvolutions as matrix multiplication: Version 2
Convolutional operators in deep networks
The problem with fully connected networks
全连接层的问题：
- 参数太多
- 没有平移不变性
How convolutions “simplify” deep networks
卷积网络的特性：
- 参数共享
- 局部连接
Advantages of convolutions
卷积网络的优势：
- 减少参数量
- 能捕捉到一些平移不变的特征
Convolutions in deep networks
书上更详细一些，建议看书
Elements of practical convolutions
1. Padding
只有奇数大小的kernel 才能 通过 padding，使得输入输出图片大小相同。padding数值为kernel大小减1，kernel_size - 1 就是偶数，然后左右或者上下各padding 一半
2. Strided Convolutions / Pooling
Grouped Convolutions
Dilations
Differentiating convolutions
What is needed to differentiate convolution?
Refresher on differentiating matrix multiplication
Convolutions as matrix multiplication: Version 1
The adjoint of a convolution
Convolutions as matrix multiplication: Version 2
卷积计算的两种方式：
为了使得计算图中的内存占用少，所以推荐第二种