A model can propose labels for unlabelled images, reducing some annotation effort. Those proposals still originate from the model and can carry its existing errors forward.
Scikit-learn describes self-training approaches that use predicted labels in semi-supervised learning. Scikit-learn — Semi-supervised learning.
Preserve label provenance
For a character-image dataset, distinguish independently reviewed labels from model-generated proposals. A later reviewer should be able to identify where a category assignment came from.
Evaluate the feedback loop
Inspect whether mistaken predictions become more common after the proposed labels enter training. Keep an appropriate independent evaluation set. A large automatically labelled dataset is not automatically stronger evidence than a smaller carefully reviewed one, particularly when unusual product details are underrepresented.
Source reviewed 25 October 2026. Interpretation and original diagrams: PLUSHIP / Qunhe. Examples are illustrative.
Related reading
A Pipeline Keeps Learned Preprocessing With the Model
An Unusual Image Is Not Automatically a Defective Product