Hands-On Machine Learning for Algorithmic Trading

The convolution operation

While scanning the input, the kernel is convolved with each input segment covered by its receptive field. The convolution operation used in CNN corresponds to the dot product between the filter and the target input area after both have been reshaped to vectors.

Each convolution results in a single number, and the result of the entire scan is a feature map that indicates activated input regions.

The following diagram illustrates the result of a scan of a 5 x 5 input, using a 3 x 3 filter, and how the activation in the upper-right corner of the feature map results from the dot product of the flattened input region and the kernel:

The most important aspect is that the filter values are the parameter of the convolutional ...