Module 04 Summary

The idea this module was built around

Depth without non-linearity is width you paid for and cannot use. Two linear layers compose into one.

What you can now do

• Prove the collapse of stacked linear layers, and fix it with an activation
• Choose a head and a loss from the target type
• Standardise features and one-hot unordered categories
• Compare a network against two baselines before claiming it helps

The trap this module removed

Reporting a network's accuracy with no baseline. On imbalanced data a constant predictor can score 0.9. The majority-class figure and a logistic regression are what make the network's number mean anything.

Figures worth remembering

Multiclass probabilities sum to 1.0 per row. Scaled features substantially beat unscaled ones with everything else held constant.

Before you move on

Add the two baselines to a model comparison you have already run. The gaps are usually smaller than expected.

Mark complete