Hands-On Machine Learning for Algorithmic Trading

The convolution stage – detecting local features

The first stage successively applies a filter, also called kernel, to overlapping patches of the input image. The filter is a matrix of much smaller size than the input, so its receptive field is limited to a few contiguous pixels of the input. Hence, it focuses on local patterns, and reduces the number of parameters and computation relative to a fully-connected layer.

A complete convolutional layer has several feature maps organized in depth slices, so that multiple features can be extracted at each location.