Convolutional Neural Network-Pooling Operation

Convolutional Neural Network-Pooling Operation

Figure 1 LeNet network structure

There is also an important operation in convolutional neural networks- Pooling . There are different types of pooling operations, including: maximum pooling , average pooling, and so on. In practice, Max Pooling has a better effect. The operation process of maximum pooling is shown in Figure 2:

Figure 2 Maximum pooling

 Before using pooling, we need to artificially determine the size and stride of the pooling window. In Figure 2, the pooling window has a size of 2x2 and a stride of 2. In the case of maximum pooling, the pooling window will acquire/retain the largest element from the feature map. The pooling operation reduces the dimensionality of the feature map while also retaining the most important information. If it is average pooling, the elements in the pooling window are averaged.

The result of feature map pooling is shown in Figure 3:

Figure 3 Pooling feature map

The role of pooling:

1. The dimension of the feature map is made smaller, and the size of the feature map is easier to manage.
2. Reduce the number of parameters and calculations in the network, and prevent network overfitting to a certain extent.
3. Make the network invariant to small deformations, distortions and translations in the input image.