"carefully engineered their soft labels" is the same thing as training the network. Just because you encode information outside of the weights doesn't mean you'

"carefully engineered their soft labels" is the same thing as training the network. Just because you encode information outside of the weights doesn't mean you're not encoding information from training data. It's like saying here's the ideal partitioning scheme, memorize this.