Google Research has published a blog post on using transfer learning for genomic prediction, with a focus on underrepresented populations. The central idea is that genomic prediction models trained mostly on one group may not perform well when applied to another group, especially when data for that group are limited.
The post describes transfer learning as a way to adapt models to such settings by making use of knowledge from populations with larger genomic datasets. This could reduce the need for large, population-specific training sets and make prediction tools more broadly applicable.
This article draws on the single source listed above. Since only one source was available, there are no differing claims to compare or reconcile.