😁Lecture 5

● Convolutional Neural Networks

Q：为什么用convolutions：

A：parameter sharing & spasity of connections

若全用fully Connected ； weight matric 会很大

// CNN 也good at caputring translation invariance

○ Convolution layer

○ Spatial dimensions, filters, padding, stride, filter size, pooling

Meaning

Why need

Padding：若无，match spatial resolution of output and input

‘

Pooling

有参数，filter；Stride了；；但是没有需要学习的超级参数；；fixed 运算， 梯度下降不会改变任何

注意：

一般f is oodd-》the this type of same convoultion gives a natural padding

We can pad the same dimension all around, 而不是左边或右边多pad这样的asymmetric padding

-》f为odd，有一个central pixel， can tell the position of the filter

Valide convolution： p=0

Same convolution：pad as much as you need to make sure the output has the same dimension as the input

Stride：