Hi Arthur,

thanks for the quick response! I thought about your answer and I wonder:

Why do we force the model to make definitive statements? Sure, with classification the target is always an one-hot encoded vector, so it would make sense to output vectors “close” to one-hot encoding. If the model is right, it will be completely right, but if the model is wrong, it will be completely wrong.

Suppose, I am randomly thinking of a number between one and two. The model has to guess it. Would a good model answer with definitive statements or with 50% 1, 50% 2?