Convolutional neural network -- understanding of pooling

Link to the original text ：https://www.cnblogs.com/booturbo/archive/2020/04/13/12693858.html

When training convolution neural network model , I often meet max pooling  and average pooling, In recent years, most image classification models have adopted max pooling, Why are they all used max pooling, What are its advantages ？

In general ,max pooling Is much better , although max pooling  and  average pooling  All the data sampling, But I feel max pooling It's more like making feature selection , The features with higher classification recognition are selected , It provides nonlinearity , According to relevant theories , The error of feature extraction mainly comes from two aspects ：（1） The variance of the estimated value increases due to the limited size of the neighborhood ;（2） The error of convolution layer parameters causes the shift of estimated mean value . Generally speaking ,average pooling  Can reduce the first kind of error , Keep more background information of the image ,max pooling  Can reduce the second error , Preserve more texture information .average pooling  More emphasis on the overall feature information sampling, It contributes more to reducing parameter dimensions , It is more reflected in the complete transmission of information , In a large and representative model , such as DenseNet Most of the connections between modules in the are  average pooling, While reducing dimensions , More favorable information is transferred to the next module for feature extraction .

average pooling  It is also widely used in the global average pooling operation , stay ResNet and Inception Average pooling is used in the last layer of the structure . sometimes , Using global average pooling near the end of the model classifier can also replace flatten operation , Make the input data into a one-dimensional vector .

 

max pooling  and  average pooling  The performance of is very helpful for designing convolution network model , Although the pooling operation has little effect on the overall accuracy improvement , But in reducing parameters and dimensions , Control over fitting and improve model performance , The role of saving computing power is still obvious , Therefore, pooling operation is an indispensable part of convolutional neural network design .