Pooling Operations in Deep Learning: From “Invariable” to “Variable”
Table 1
Summary of basic pooling operation.
Pooling method
Characteristic
Sketch map
Max pooling
The max pooling operation is simple, but the features with strong influencing factors are ignored and lost information
Maximum two-mean pooling
Maximum two-mean pooling operation improves the disadvantage of max pooling to a certain extent, which is ignoring features with larger influencing factors. But there is some problem of loss information
-max pooling
-max pooling operation preserves more information than max pooling and preserves the relative order of eigenvalues, but there is no absolute position information
Average pooling
Average pooling operation takes global information into account, no information is lost, and overfitting is reduced. However, the features tend to be smooth, and the information of prominent features cannot be extracted
Median pooling
Median pooling operation can learn the characteristics of edge and texture structure and has a strong antinoise ability
Sum pooling
Sum pooling operation considers global information. There is no information loss, but it is vulnerable to extreme information
