Fig. 3: Sample size estimation with inflection points.

The top and middle rows are the results of individual datasets, while the bottom row is the combination of all eight tested datasets. The black striped line represents the autoencoder loss at the point of the inflection. The shaded region represents the error bars, with error determined as the autoencoder loss at ± ln(n), where n is the sample size at which the inflection point of the autoencoder loss occurs. The points are shaded by sample size. For each of these datasets, the autoencoder loss method appears to provide an unbiased estimate of the minimum convergence sample. The top row demonstrates appropriately sampled data while the middle row shows statistical power estimation on oversampled and undersampled data. The bottom row shows that linear interpolation using auto-encoder loss function generally works well in estimating learnability. a Test Area-Under-the-Curve Metric as a function of autoencoder loss. b Test Area-Under-the-Curve Metric as a function of the derivative of the autoencoder loss. c Test Area-Under-the-Curve metric as a function of the double derivative of the autoencoder loss. d Linear interpolation of autoencoder loss with respect to Test Area-Under-the-Curve metric.