🍎Lecture 11

● CNN architectures

○ AlexNet, VGG, ResNet

//why choose filter of certain size

//why should we have so many layers

VGG：make filter 3*3， advantage：reset filed 7*7

Advantage1

Advantage2:

可以learn more complex的

Motivation of resnet

Purpose of ResNet：

• VGG：19layers（（（ps GoogLeNet有22layers

Visual geometry group

Q：why 3*3 filters

A：it would be the smallest odd number that look at some special content beyond that central pixel

ResNet：

Deeper 不一定越好，has optmization issue

-》

//（skip-》gradient flow improve）make opt easier：因为output 的gradient get straight pass back；；this skip connection，让output 与input 的connection much shallower

//global pooling layer 和no FC at the end（only FC 1000 to ouput classes

○ Receptive field

Means：the size of the input patch

CNN中：size of input

C个channel；C个filter

//more 参数-》更容易overfit

&有了更多non-linear

/////

C'filter，Cchannel-》c squeare

FLOP：每秒浮点计算（floating-point operations per second）

///

AlexNet： each is a single layer

VGG（deeper network）： stage；；better performance

‘