Entropy is a heuristic that measures the model’s uncertainty about the class of the image. The higher is entropy – the more “confused” the model is. As a result, data samples with higher entropy are ranked higher and offered for annotation.
Below is a mathematical formula for entropy calculation:
Source
Imagine that our model predicts whether an object is a dog or a muffin. Let’s say it gave the following probabilities for instances A and B:
- Instance A: “dog” – 0.5, “muffin” – 0.5.
Entropy = H([0.5,0.5]) = 1.0 - Instance B: “dog” – 1.0, “muffin” – 0.0.
Entropy = H([1.0,0.0]) = 0.0
As you see, the model was completely confused when the class probabilities were equal. In contrast, when one of the classes had a 100% probability, the entropy was equal to 0 since there was no uncertainty.
In case you are wondering how dogs might look similar to muffins. They really do!
Entropy might be useful in datasets with very diverse images and classes.
Learn more about the other heuristics: