Module 01 Summary

The idea this module was built around

Depth buys learned intermediate representations. That is the entire advantage, and it only pays when the problem is non-linear.

What you can now do

• Check what hardware you have and write device-agnostic code
• Show why a linear model fails on XOR and one hidden layer does not
• Demonstrate that capacity without data is memorisation
• Seed a run so someone else can reproduce it

The trap this module removed

A model that reports low training loss on pure noise has memorised, not learned. With 512 hidden units and 40 training rows there are more parameters than examples, and training loss stops being evidence of anything.

Figures worth remembering

Linear on XOR: 0.5. One hidden layer: 1.0. Same data, same optimiser — only the representation changed.

Before you move on

Take a result you reported recently and ask whether someone could reproduce it from what you wrote down. If not, the missing piece is usually the seed.

Mark complete