The interactive film is in the book's Companion

Play it, pause it, turn it and drag it with Mathematics for AI and Machine Learning.

Get the book

Temperature Scaling: Calibrating a Softmax

Chapter 13 · Mathematics for AI and Machine Learning · Advanced · 1:25

An interactive live film: logits divided by T move the softmax point across the probability triangle, freezing to the argmax corner as T → 0 and melting to the center as T → ∞, with the argmax unchanged. Tune T so the top confidence matches the 52% validation accuracy.

Narration

A classifier produces raw unnormalized scores called logits for each competing class.

Dividing logits by temperature and exponentiating maps them onto the probability simplex, where all class probabilities sum to one.

The rest of the narration is in the book's Companion.

More from Chapter 13

Interactive1:35Normalizing Flows: Chaining Invertible MapsChapter 13Preview1:38Independence as an Outer ProductChapter 13Preview1:40Conditional Gaussian: A Slice of the BellChapter 13Preview1:22Random Fourier FeaturesChapter 13Preview1:46Change of Variables: Same Probability, Different AreaChapter 13Preview1:35The von Mises-Fisher Distribution: A Gaussian for DirectionsChapter 13Preview1:42The Dirichlet Distribution: A Density on the SimplexChapter 13Preview1:14Gaussian Mixture Model: Multimodal TerrainChapter 13Preview1:41Whitening: PCA and ZCAChapter 13Preview1:36Beta and Bernoulli: Updating with Pseudo-CountsChapter 13Preview1:43ROC Curve and AUC: the Mass Above the DiagonalChapter 13Preview1:40Hoeffding's Inequality: How Fast a Sample Mean SettlesChapter 13Preview1:34Mahalanobis Distance: The Ellipsoids of a GaussianChapter 13Preview2:09Random Directions Are Nearly OrthogonalChapter 13