Hands-On Machine Learning for Algorithmic Trading

The pooling stage – downsampling the feature maps

The last stage of the convolutional layer downsamples the input representation learned by the feature map, to reduce its dimensionality and prevent overfitting, lower the computational cost, and enable basic translation invariance. The precise location of the learned features is not only less important for identifying a pattern or object, it can even be harmful, because the locations will likely vary for different instances of the target. Pooling lowers the spatial resolution of the feature map as a simple way to render the location information less precise.

It is optional, however, and many architectures do not use pooling, or only use it for some convolutional layers.