02·Loss function — measuring how wrong the network is

Introduction

Training is optimisation — reducing a number called the "loss". But what exactly is that number, why must it be differentiable, and why does the choice of MSE vs cross-entropy decide whether the network learns at all? This lesson lays the foundation: what a loss function is, how it differs from a metric, the main loss families for classification and regression, and which numerical pitfalls lurk in practice.