[apstroll]

Under a crossentropy loss the output activations do absolutely represent a probability distribution, since that is what we're modeling.