The average loss is not really helpful for my models, because few extremely wrong updates, will pull up the likelihood to very large values (high stochasticity).
I am now plotting loss curves, but these are not helpful with the large extremes. This metric should be improved, because it is a great diagnostic to see if the chains have converged.
The average loss is not really helpful for my models, because few extremely wrong updates, will pull up the likelihood to very large values (high stochasticity). I am now plotting loss curves, but these are not helpful with the large extremes. This metric should be improved, because it is a great diagnostic to see if the chains have converged.