Introduction
Ordinary code usually crashes when something is wrong. A neural network with a bug often runs perfectly happily and just learns badly: a few percent worse than it should, or not at all. A wrong normalization, labels shifted by one position, images loaded upside down, a loss function that doesn't match the output layer: none of these raise an error.
That's why experienced practitioners never start with a long training run. They first spend ten minutes on a few sanity checks that rule out most bugs.
Think of a pilot before take-off. They don't just start the engines and hope. They walk around the plane, check the instruments read sensible values, and test the controls on the ground. Each check is quick and boring, and each one prevents a disaster. This topic is the pre-flight checklist for training a network.