Lecture 5: Convolutional Neural Networks

CS231n

Lecture 5: Convolutional Neural Networks

History

Rosenblatt Perceptron $\to$ Rumelhart Back-propagation $\to$ Hinton RBM $\to$ DBN $\to$ AlexNet
Hubel & Wiesel $\to$ FUKUSHIMA Neurocognition $\to$ LeNet $\to$ AlexNet
现如今CNN已经得到了广泛的应用

CNN

实现的是特征的层级抽象

Conv层

提取特征
要素：$pad,w,stride$

Pool层

简化特征
max-pooling等

FC层

一般作为分类器

Summary

• Trend towards smaller filters and deeper architectures, getting rid of POOL/FC layers (just CONV)
• [(CONV-RELU)*N-POOL?]*M-(FC-RELU)*K,SOFTMAX, but recent advances such as ResNet/GoogLeNet challenge this paradig