Reference · Glossary

Probability for ML

Last updated

Classification with uncertain outcomes, calibrating spam filters, choosing decision thresholds, and understanding softmax outputs from LLM token predictions.

#When to use

Classification with uncertain outcomes, calibrating spam filters, choosing decision thresholds, and understanding softmax outputs from LLM token predictions.

#When not to

Deterministic lookups where probability adds no value — e.g. fetching a user record by exact ID.

#Example

Model output: P(spam | email) = 0.91
Threshold 0.5 → label "spam"
Threshold 0.95 → label "not spam" (fewer false positives)

Probability connects **predictions**, **thresholds**, and **loss functions** in one framework.