Figure 2:

Example workflow for human-in-the-loop training using MNIS T²⁰. a) Abstraction of steps in a human-in-the-loop learning procedure. b) Example MNIST digits predicted as “3” after a few iterations. The misclassified digits can be corrected by assigning them to the right label (“8” and “7”). c) Example of correctly predicted digits. Images with easy patterns such as “1”s are mostly correctly identified after only a small number of manually labeled examples. d) Researchers can inspect the confusion matrix available in Piximi to see which classes are misclassified the most, to focus their efforts where most needed.