Lesson 45 of 48
Backpropagation from scratch
The chain rule, applied at scale, in your own code.
This is the lesson the previous 44 were for. When your own NumPy network crosses 90% on held-out digits with no framework imported anywhere, you will have earned the claim that you know how this works.
Do this
Train a two-layer network on MNIST to at least 90% held-out accuracy, in NumPy, with no machine-learning framework imported anywhere in the file.
The question that unlocks the next lesson
What is backpropagation, mathematically?
- AA search over random weight settings
- BThe chain rule applied in reverse through the computation graph, reusing shared intermediate results
- CA way of initialising weights
- DA method for normalising inputs