The forward and backward pass built from nothing, then stacked until it is deep.
10h · 12 lessons · 1 challenge
Before this: Numerical Computing & Stability, Neural Networks: the First Pass
The second pass at neural networks, and the one that takes the backward pass apart. Logistic regression is rebuilt as a one-neuron network so that the computation graph and its derivatives are visible, then vectorized across a batch, then widened into a hidden layer, then deepened.
You implement all of it in NumPy before touching a framework, because the shapes are the whole difficulty and a framework hides them.
Annotate every tensor in a deep network with its shape, against the clock. One wrong dimension ends the run.
Worth 450 XP.
