Skip to main content

4 · Optimisation

How a model is actually fitted when no formula exists — choosing the loss that defines the answer, following the gradient downhill, setting the one learning-rate knob whose failure modes are computable, and deciding how many rows each update is allowed to see.