Binary Cross-Entropy
Binary cross-entropy is used for binary classification when the model outputs a probability for class 1.
Let:
y = target, either 0 or 1
p = predicted probability of class 1
The binary cross-entropy loss is:
loss = -[y log(p) + (1 - y) log(1 - p)]
For this chapter, focus on the two cases.
If the true target is 1:
loss = -log(p)
If the true target is 0:
loss = -log(1 - p)
The loss is small when the model assigns high probability to the correct class.
Small numerical check
If the target is 1 and the model predicts p = 0.8, then:
loss = -log(0.8)
This is smaller than:
-log(0.2)
because predicting 0.8 for the true class is better than predicting 0.2.
The target is 1. Which predicted probability gives smaller binary cross-entropy: 0.9 or 0.2? Enter the smaller-loss probability.
Compute it first, then check your number.
HintTrue class is 1
For target 1, the loss is -log(p).
SolutionWork it out
Since the target is 1, the model should assign high probability to class
0.9gives a smaller loss than0.2.
The target is 0, and the model predicts probability p = 0.1 for class 1. What probability did the model assign to the correct class 0?
Compute it first, then check your number.
HintBinary probabilities sum to one
Class 0 probability is 1 - p.
SolutionWork it out
The model assigns p = 0.1 to class 1, so it assigns 1 - 0.1 = 0.9 to
class 0.