Autograd
The same multilayer perceptron (MLP), but now trained with automatic differentiation — no more hand-written backward pass.
- 2.1 What Changes
- 2.2 The Value Class: Wrapping Numbers
- 2.3 Recording Operations: Add and Multiply
- 2.4 More Operations and Syntactic Sugar
- 2.5 Backward: Walking the Graph
- 2.6 Parameters as Values
- 2.7 The New Training Loop
- 2.8 Same Results, Less Code
The big idea: Every operation records how it was computed. The chain rule can then be applied automatically by walking backward through the computation graph.