Module 06 Summary

The idea this module was built around

A convolution applies one filter at every position, so a pattern learned in one corner is recognised in the other. That is translation equivariance, and it is why CNNs need far less data than dense networks for images.

What you can now do

• Explain what flattening throws away
• Apply the output size rule and verify it in code
• Measure the flatten width instead of guessing it
• Evaluate an imbalanced problem with per-class recall

The trap this module removed

Accuracy on imbalanced image data. A model can score respectably overall while a class has zero recall — every example of it misclassified, with nothing in the accuracy figure to say so.

Figures worth remembering

Kernel 3 with padding 1 preserves size; stride 2 halves it. Parameter count depends on channels and kernel, never on image size.

Before you move on

Rank a model's errors by confidence and open the top three images. In real datasets a surprising share are labelling mistakes.

Mark complete