Alex H. Williams

Bits per Spike as a Betting Game

A simple re-scaling of the heldout log-likelihood yields an intuitive metric of model performance in units of recording time which answers “how much heldout data would I need, on average, to disprove the null.”