Hi Andreas, thanks for reading!

1. Yes, you are correct, thanks for pointing that out.
2. In many cases, yes. However, it depends on whether or not you chose to run a softmax or other activation function over the last output layer of the network. If you chose to use some linear function, there is a high chance the output classification won’t sum to 1.0 and will be some other metric determined by the model (that can be turned into probabilities via softmax.