Fig. 2: The DeepDive approach (blue) can reduce errors and improve R2 relative to SQS (red) under different bias conditions.
From: DeepDive: estimating global biodiversity patterns through time using deep learning

Accuracy of biodiversity estimations relative to simulated diversity for test datasets where (A) shows estimates made using DeepDive and (B) SQS at quorum level 0.6, in both cases the black line of slope 1 indicates the goal of these methods to make as close to a 1:1 estimate as possible. The variation in R2 and relative MSE where (C) shows estimates on a test set generated under the same parameters as the training set, and for test sets generated to under different parameterisation to represent conditions of strong (D) temporal, (E) taxonomic and (F) spatial bias and for patterns that are rare in the training simulations (G) mass speciation and mass extinctions, (H) diversity dependence followed by mass extinction (see “Methods” for more details) for DeepDive and SQS. Data are presented as median values +/− the interquartile range, whiskers at 1.5 IQR. n = 100,000 (100 time bins × 1000 simulations). We note that the performance of DeepDive trained models substantially improved for settings (G) and (H) after re-training the model including sudden changes in speciation and extinction rates and diversity dependent processes in the simulations (Supplementary Fig. 7).