Global average Pooling

Motivation

CNN has too many filters because CNN stack convolution layers. It means CNN has too many feature maps, parameters. It makes over fitting the network. So we need to decrease parameters.

And CNN + FC has too many parameters, and make disappear information of position.

 

Main Idea

Average pooling

Let's see average pooling. max average makes x([1,4,4,3]) to [1,2,2,3] with pool_size=2, strides=2.

Max pooling makes similar result. Just not average, max.

We know pooling layer decrease feature size.(=feature maps, parameters, dimension)

 

GAP(Global Average Pooling)

GAP decrease feature size rapidly. Difference between pooling and GAP is in purpose. purpose of GAP is making feature to 1-d. 

GAP makes size of the features [height, width, channels] to [channel,] and remove or decrease FC layers.

GAP can be used with FC layers, to make feature to inputs of FC layers.

We can impelment GAP simply using Average Pooling