Machine Learning with Swift

Convolutions in CNNs

The first layer applies the filters to the input image, and sends the resulting feature matrices downstream. The next layers apply their filters to the input they get, extracting features from a higher and higher level (see Figure 9.6 - 9.8). In the task of image classification (or object recognition), CNNs gradually adjust the filters during the training to extract useful features from the picture. In this case, useful features are different patterns, like eyes, beaks, wheels, and so on:

Figure 9.6: Each CNN layer learns a set of convolutional kernels. Here you see outputs of the different filters from the 2nd convolutional ...