Chapter 15
Backpropagation
Computing Gradients Through Computational Circuits
Backpropagation is the algorithm that makes training neural networks tractable. By viewing computation as a circuit, we can compute all partial derivatives in two efficient passes — one forward, one backward.
7 sections~53 minutes
Sections
1
15.1
The Gradient Problem
8 min · 5 quiz questions
2
15.2
Circuit Model — Nodes, Wires, Functions
7 min · 4 quiz questions
3
15.3
The Forward Pass
6 min · 4 quiz questions
4
15.4
Cutting a Wire
7 min · 4 quiz questions
5
15.5
The Function Node Rule
10 min · 5 quiz questions · Interactive diagram
6
15.6
The Splitter Node Rule
7 min · 5 quiz questions
7
15.7
Putting It All Together
8 min · 5 quiz questions