Skip to content
The loss curve

Cross-entropy

The standard loss for classification: −Σ y·log(p). Designed to pair cleanly with softmax/sigmoid outputs.

Companion explanation in Step by Token, chapter 6.

Where this term gets built