A factory could start with an image model pretrained on a broad collection of objects and replace its final classifier with a head for “defective” and “acceptable.” The model’s feature extractor supplies a useful starting representation for the factory’s labeled photographs.
In one approach, freeze the extractor and train only the new head. In another, also update some pretrained layers: this is fine-tuning. Both are transfer learning. Freezing means the extractor’s weights stay fixed during the head’s training; it does not mean the whole pipeline has no trainable parameters.
Reuse can reduce the training work and labeled data needed, but the benefit depends on how well the source representation fits the target. Compare held-out results and avoid assuming general object features capture every manufacturing defect. Transfer learning describes reuse across a task or domain boundary, not a guarantee of better accuracy. Zero-shot learning pushes reuse further: the model handles target categories it saw no labeled examples for.
Sources
- PyTorch: Transfer Learning for Computer Vision — Shows fine-tuning a pretrained network and freezing it as a fixed feature extractor.
Go deeper
- Keras: Transfer learning and fine-tuning docs
Work through layer freezing, a new classifier and careful unfreezing for adaptation.