Training with domain adaptation greatly reduces the number of positive tests, particularly between geographically isolated populations. Note that other model violations are likely present too which we did not investigate with our simulations.

(6/7)

We then applied this approach to detecting introgression between ABC Island brown bears and other populations. ABC Island bears are know to have had introgression with polar bears and this has been shown to cause false positive tests for introgression between them other populations.

(5/7)

After training a CNN with domain adaption using a small number of datasets simulated with ghost introgression, the CNN performed nearly as well with datasets that had ghost introgression (orange ROC curves + confusion matrix) as it performed on datasets without (blue ROC curves).

(4/7)

We trained a CNN to detect introgression between two populations. This network performed well in the absence of model violations (blue ROC curves). But it performed pretty poorly when applied to datasets with ghost introgression from a third population (orange ROC curves + confusion matrix).

(3/7)