perClass Mira provides an active learning tool that helps us to understand what examples the model did not see in training.


You may enable it using the Show unknown toolbar button or by pressing u (unknown)



There is an extra decision "Unseen in training" with high transparency. The extra decision highlights the pixels that the classification model considers very different from anything labeled. 


For example, the leaf on the right side is largely rejected being slightly different than the two leaves we labeled.


We may add extra labeling and retrain the model using the Retrain toolbar button




The Show unknown tool allows us to label in the areas needed and, thereby, building representative training sets.


TIP: It is generally better in perClass Mira to define less but accurate and representative labels. Use small brush with at least 2 pixels (to allow assign-stroke-to-class)