I'm reading a machine learning paper, and in it, there is train and test NLL and PPL.
I know that NLL stands for "negative log-likelihood". But I don't know about PPL. What does it mean?
PPL stands for perplexity. This quantity is directly computable from the negative log-likelihood (and the number of samples), so it's odd that the authors would report both. A given sub-discipline tends to prefer one or the other.