本文共 1521 字，大约阅读时间需要 5 分钟。

吴恩达 Andrew Ng

深度卷积网络：实例探究

LeNet - 5

• 
• Conv → Pool → Conv → Pool → FC → FC → Output

AlexNet

• 
• Multiple GPUs, Local Response Normalization

VGG - 16

• 
• 一共16 层（不算池化层）

ResNet (残差网络)

Residual block

• shortcut/skip connection (跳远连接)
• 

Residual Network

• 
• 信息传递到更深层的网络
• 当网络层数达到一定的数目以后，网络的性能就会饱和，再增加网络的性能就会开始退化，但是这种退化并不是由过拟合引起的，因为实验发现训练精度和测试精度都在下降，这说明当网络变得很深以后，深度网络就变得难以训练了
• 有利于解决梯度消失和梯度爆炸 (vanish and explode gradient)
• 多用于训练深度网络

Why do residual networks work?

• 残差块学习恒等函数 (identity function) 很容易，不会降低网络效率
• same convolution ($z^{[l+2]}+a^{[l]}$ make sense)

1×1 convolution

• the number of channel
• Network in Network
• 降低信道数量 (shrink $n_c$)，或者增加、不变
• 给网络添加了一个非线性函数

Inception Network

• 
  
• 由网络决定使用多大的过滤器，怎样的组合
• computational cost
• bottleneck layer 瓶颈层
  
• Szegedy et al., 2014, Going Deeper with Convolutions
• GooLeNet

Transfer Learning 迁移学习

• use someone else’s pre-trained weighs
• freeze all of the earlier layers’ weight
• more data, the smaller layers you freeze

Data Augmentation

• Morroring operation (镜像对称)
• Random Cropping (随机裁剪)
• Color Shifting (颜色变换)
• PCA Color Augmentation

The State of Computer Vision

• less data, more hand-engineering
• lots of data, simple algorithms and little hand-engineering
• Two sources of knowledge: Labeled data, Hand engineered features/network architecture/other components

• Tips

  • Train several networks independently and average their outputs
  • Run classifier on multiple versions of test images and average results (10-crop)
  • Use architectures of networks published in the literature, open source implementations, pretrained models

  ​