Engineering Sandbox rollout · bundle-mounted ML essays
Can the class change without the probability changing?
Keep the temperature fixed and move the decision threshold past its predicted probability. The class flips because the cutoff changes, not because the model learns. A sigmoid output is a model estimate, not a guarantee of calibrated probabilities.
How to use this page
Compare two cutoffs for the same temperature, then inspect log loss. Thresholding changes the decision; log loss evaluates the probability assigned to the observed class. A confident wrong probability receives a larger penalty than an uncertain one.
Loading Logistic Regression lesson…