I'm reading the AlexNet paper. In section 4, where the authors explain how they prevent overfitting, they mention
Although the 1000 classes of ILSVRC make each training example impose 10 bits of constraint on the mapping from image to label".
What does this mean?