1. Pooling Layers
2. The pooling layer is used to reduce the dimension to speed up the operation
3. Max Pooling
Divide the image before dimension reduction equally, and fill the corresponding area after dimension reduction with the maximum pixel value of the corresponding area before dimension reduction
Equivalent to convolution kernel size f = 2, stride s = 2
The idea of Max Pooling: If a feature is detected in this area, then keep the maximum value of this area, if no feature is detected in this area, then the maximum value of this area will not be very large
4. The pooling layer has fixed hyperparameters (such as f, s) but no parameters w, b that need to be learned
5. Another example of pooling layer
6. For the input of multiple channels, perform the Max Pooling operation independently on each channel
7. Average Pooling (average pooling, not commonly used)
That is to replace the maximum value operation of Max Pooling with the average value operation
Only in very deep neural networks, the dimensionality drops rapidly (eg from 7 * 7 * 1000 to 1 * 1 * 1000)
8. Hyperparameters of the pooling layer
1) f is the size of the convolution kernel
2) s is the moving span
E.g,
3) Use Max Pooling or Average Pooling
4) Usually Padding is not used in the pooling layer, i.e. p = 0
But there are exceptions