9 · Evaluation, Metrics & Diagnostics · Calibration & Decision Thresholds

Temperature scaling

Common in neural networks: a single temperature rescales logits, preserving class ranking while changing confidence. The important practical question is not only how the technique is defined, but what assumptions it introduces, which data are allowed to influence it, and how its effect should be validated on unseen evidence.

Reference lessonPython exampleVisual explanation

Intuition first

What this concept means in practice

The practical value of Temperature scaling comes from understanding both the transformation and the boundary around it: what information is allowed to enter, what assumption is being made, and how you know the result is still valid after the transformation.

A beginner-friendly way to reason about it is to start with a tiny case where the correct result can be checked independently. Once the mechanism is clear, scale the exact same reasoning to larger tables, pipelines or models.

PurposeUse for multiclass neural networks when ranking is good but confidence is too sharp/flat.

MechanismFor neural-network logits, learn a single positive temperature that rescales confidence without changing class ranking.

EvidenceInspect intermediate and final output; compare with an independent expectation.

Main cautionIt is a limited correction and cannot fix arbitrary class-specific calibration errors.

Mechanism

Trace the operation from input to decision

For neural-network logits, learn a single positive temperature that rescales confidence without changing class ranking.

1Input→

2Apply rule→

3Inspect state→

4Validate→

5Use result

Key rule

softmax(logits/T)

Visual explanation

Make the structure visible

The interactive view uses a concept-specific plot when the topic maps naturally to one; otherwise it uses a workflow view instead of leaving a broken placeholder.

Loading visual…

Practical example

Where you would use it

Divide logits by T>1 to soften an overconfident classifier before softmax.

Use when

Use for multiclass neural networks when ranking is good but confidence is too sharp/flat.

Pitfall

What can make the result misleading

Watch out

It is a limited correction and cannot fix arbitrary class-specific calibration errors.

A useful diagnostic question is: Could the same code still run successfully if the analytical assumption were wrong? If yes, add an explicit validation check rather than relying on execution success.

Implementation

Miniature Python example

Keep the example small enough that you can inspect each stage manually.

Python

# Purpose: demonstrate Temperature scaling with a small, inspectable example.
# Follow the comments and printed stages to connect each operation with its result.
# Import the library or helper used in this example.
# Step 1 — Import the module so its functions/classes are available to the rest of this example.
import pandas as pd
# Import the library or helper used in this example.
# Step 2 — Import only the named objects needed by the following steps, keeping dependencies explicit.
from sklearn.preprocessing import StandardScaler, RobustScaler

# Create a small labelled dataset that is easy to inspect by eye.
# Step 3 — Construct `X` as a tabular object with named columns for inspectable analysis.
X = pd.DataFrame({"income":[40,45,48,52,56,60,64,70,78,120],"age":[22,25,28,31,34,37,40,43,46,49]})
# Print this intermediate result so you can verify the workflow step by step.
# Step 4 — Display the current value explicitly so the result/state can be inspected during execution.
print("STEP 1 · Raw means:", X.mean().round(2).to_dict())
# Learn the transformation from this data and apply it in one step.
# Step 5 — Fit the transformation on the training input and immediately transform that same input.
std = StandardScaler().fit_transform(X)
# Step 6 — Fit the transformation on the training input and immediately transform that same input.
rob = RobustScaler().fit_transform(X)
# Print this intermediate result so you can verify the workflow step by step.
# Step 7 — Display the current value explicitly so the result/state can be inspected during execution.
print("STEP 2 · Standard-scaled first row:", std[0].round(2).tolist())
# Print this intermediate result so you can verify the workflow step by step.
# Step 8 — Display the current value explicitly so the result/state can be inspected during execution.
print("STEP 3 · Robust-scaled outlier row:", rob[-1].round(2).tolist())

Expected / illustrative output

STEP 1 · Raw means: {'income': 63.3, 'age': 35.5}
STEP 2 · Standard-scaled first row: [-1.07, -1.57]
STEP 3 · Robust-scaled outlier row: [3.18, 1.0]

Implementation checklist

Before you move on

Can you state what data or object enters the operation?
Can you explain what changes and what must remain invariant?
Have you checked the result on a tiny case you can verify independently?
Have you considered the main failure mode described above?
Can the operation be reproduced from code/formulas and documented assumptions?