Determine the cause of the public-adaptation performance plateau

Determine the dominant cause of the performance plateau observed when expanding the public chalcogenide RHEED adaptation training set beyond approximately 100 images under the fixed partition and update budget.

Background

The public-image adaptation experiment found that relatively small training sets captured much of the improvement in overall Kikuchi localization, while enlarging the training set from 100 to 244 images produced only a small change in IoU under the fixed partition and update budget. A separate presentation-control experiment also showed only a modest improvement when the 244-image run received more updates. The available results therefore do not resolve whether the apparent plateau is caused primarily by limited optimization, the number and composition of training images, checkpoint-selection effects, or other factors within the experimental design.