Evaluation
Cross Entropy Loss
Compare loss behavior as predicted probabilities change.
Loss Curve
Higher probability on the true class lowers loss.
Binary Classification
True label: Class 1 (Positive)
where y is the true label and p is the predicted probability
Prediction probability
Class 0
50%
Class 1
50%
Loss
0.693
loss = 0.693
-log(p_true)
-log(p_true)
Implementation note
Cross entropy punishes confident wrong answers sharply. That is why evaluation has to measure calibrated task success, not just fluent output.