Mastering Machine Learning Algorithms

Other useful layers

Even if convolution and pooling layers are the backbone of almost all deep convolutional networks, other layers can be helpful to manage specific situations. They are as follows:

• Padding layers: These can be employed to increase the size of a feature map (for example, to align it with another one) by surrounding it with a blank frame (n black pixels are added before and after each side).
• Upsampling layers: These increase the size of a feature map by creating larger blocks out of a single pixel. To a certain extent, they can be considered as a transformation opposite to a pooling layer, even if, in this case, the upsampling is not based on any kind of interpolation. These kinds of layers can be used to prepare the feature ...