Line Plot from Scientific Research

Distribution of description length gaps Δℋ(m)=ℋ(m)−ℋ(m*); each line corresponds to a different dataset D . The true generating model has Δℋ(m)=0 , indicated by a horizontal dashed line. Positive gaps correspond to models m that are less plausible than the true generating model m *, and vice versa. When there is a model m with negative gap for a given dataset D , the true model is unlearnable from that dataset because m * is not the most plausible model given the data.

DOI