In this video, you will learn about
When we apply 2D conv to an imagen we tend to lose pixels located on the perimeter of our image. That's where padding is useful, adding some border to extract better values. Also, padding can be apply to data (e.g. images) when dimensionality of differs. specifically, imagine two images of different sizes. Padding can make the smaller image of the same size as the other one without croping or any other way when some information would be lost.
When we deal with image datasets the computer tends to have many parameters inside. So we can apply Pooling to downsample the parameters inside our net.
"Dilated Convolutions are a type of convolution that “inflate” the kernel by inserting holes between the kernel elements. An additional parameter (dilation rate) indicates how much the kernel is widened. There are usually spaces inserted between kernel elements." Fisher Yu and Vladlen Koltun
in this blog post , Ferenc Huszár explains the advantage and usage of dialated convolutions. In a nut shell, in his words, "it allows for very large receptive fields while only growing the number of parameters logarithmically".