The convolution, the layer built from it, and the architectures that made depth work.
11h · 14 lessons · 1 challenge
Before this: Structuring Machine Learning Projects
A fully connected layer on an image has more parameters than you have images. The convolution is the answer, and this module derives it from edge detection rather than presenting it as a definition: what a filter is, what padding and stride do, and how it extends over volumes.
Then the architectures, as a sequence of problems and solutions: LeNet, AlexNet and VGG, then ResNets and why depth suddenly became possible, then Inception and the efficient families.
A feature to detect and a three-by-three grid. Hand-build the filter that finds it, before you are allowed to train one.
Worth 450 XP.
