Deep Learning with PyTorch Quick Start Guide

Pooling layers

Convolutional layers are typically stacked using pooling layers. The purpose of a pooling layer is to reduce the size, but not the depth, of the feature map generated by the preceding convolution. A pooling layer retains the RGB information but compresses the spatial information. The reason we do this is to enable kernels to focus selectively on certain nonlinear features. This means we can reduce the computational load by focusing on the parameters that have the strongest influence. Having fewer parameters also reduces the tendency to overfit.

There are three major reasons why pool layers are used to reduce dimensions of the output feature map:

• Reduces computational load by discarding irrelevant features
• Smaller number of ...