Logistic regression or neural networks with softmax activation use $ln p$ when manipulating probabilities. Normalization to a total probability of 1 is usually

Logistic regression or neural networks with softmax activation use $ln p$ when manipulating probabilities. Normalization to a total probability of 1 is usually delayed to the end (so intermediate values are actually $ln p_i + c$ where c will only be computed during normalization). This representation of probabilities has the nice property that Bayes' law turns into simple addition. The linked approach sounds very similar, calling $-ln p$ rank. However it appears to use a different normalization approach.