The result of the image after convolution and activation generally contains a lot of original image information.
In the above figure, the convolution kernel extracts continuous pixels in the vertical direction, but the enhanced pixels only account for 1/3 of the result. For the multi-layer network, the rest of the less important information is also passed in. One layer of network causes unnecessary waste, so it is necessary to use pooling to aggregate the results obtained by the base layer of the volume. The theoretical basis of pooling is that the pixels in adjacent positions of the image are related, and even if the image is sampled at a certain size, the obtained result can still retain most of the information. Commonly used pooling methods are max pooling and mean pooling. Another important role of pooling is to bring some translation, rotation, and perspective invariance to convolutional neural networks.
The figure above shows how pooling brings translation invariance to the model. A segment of continuous pixels in the vertical direction in the original image is shifted to the right by one unit to obtain the translated image. After the two are max pooled with a size of 2*2 and a stride of 1, the same result is obtained. Pooling can achieve a similar effect for rotation and perspective. However, pooling makes the model resistant to translation, rotation, and perspective. It is not without restrictions. First of all, the effect of pooling depends on the result of convolution. If the convolution does not capture the above transformed features, the pooling is also lost. The ability to resist translation, rotation, and perspective. Secondly, the resistance of pooling to the above three transformations is related to the size of the pooling. The larger the pooling size, the stronger the resistance, but the more information is lost. The advantages and disadvantages need to be Designer tradeoffs.