Fig. 3.

Building blocks of convolutional NNs and key mathematical concepts. A) Key mathematical operations executed in NNs: Fully-connected node, convolution, 2 × 2 MaxPooling (top to bottom). Note that activation is also performed for the convolution operation. B) Common activation functions and their definition. C) A simplified convolutional NN. The input image is passed to multiple (in this case: two) convolutional layers. After this, a 2 × 2 MaxPooling is performed, followed by flattening into a one-dimensional vector (which has, in this case, 50 elements). One fully-connected layer follows before the final layer returns the output (in this case: two scalars).