Temperature Scaling
Definition
Temperature scaling adjusts the probability outputs of a classifier without changing its chosen class.
Temperature scaling is a post-calibration method for classifiers: a positive scalar T is learned on separate validation data and divides the logits before softmax, without changing the predicted class.
Simply explained
The decision remains the same, but its displayed confidence is adjusted. Separate examples are used for tuning this.
Why it matters in practice
The term helps to select, develop, evaluate AI systems in a well-founded manner and realistically assess their technical risks.
Example from business
A quality model learns the temperature parameter on separate validation data and then checks the reliability diagram and ECE.
Limits and pitfalls
- Confusing temperature scaling with sampling temperature, using the same dataset for fitting and evaluation, or assuming calibration without verification under data shift and for each subgroup.
Origin & history
Guo et al. demonstrated in 2017 that simple post-processing often effectively calibrates modern neural networks.
Use in mid-sized companies
- Possible application: A quality model learns the temperature parameter on separate validation data and then checks the reliability diagram and ECE.
Frequently asked questions
What does temperature scaling mean?
Temperature scaling is a post-calibration method for classifiers: a positive scalar T is learned on separate validation data and divides the logits before softmax, without changing the predicted class. What mistakes should be avoided? Confusing temperature scaling with sampling temperature, using the same dataset for fitting and evaluation, or assuming calibration without verification under data shift and for each subgroup.
Sources
- On Calibration of Modern Neural Networks · Guo et al., On Calibration of Modern Neural Networks
Source checked on 10.10.2026. Editorial application examples are hypothetical; no evidence of real customer results.
Editorially reviewed on: