Transfer Learning

Reusing knowledge from one task or dataset to improve performance on another.

Start with a model that has already seen images

An archive has a small labeled set of framed and loose prints. It can start from an image network trained on a different collection.

Freeze the image representation

Keep the pretrained feature extractor fixed and train a new output layer on the archive’s labels. This gives a relatively inexpensive baseline.

Adapt some features

Allow selected layers to update on those labels as well. The extra flexibility can help with an unfamiliar image style or allow the model to memorize a small collection.

What would show that transfer helped?

Reserve whole acquisitions for evaluation so alternate views of one print stay together. Compare the frozen model, the adapted model and a suitable baseline on those same held-out acquisitions. Report where transfer helps and where frames or mounting boards still cause errors.

Common Forms of Transfer Learning

Transfer learning often appears through fine-tuning, feature reuse, adapter methods, or parameter-efficient techniques such as LoRA. The broader principle is simple: reuse general knowledge, then specialize only where needed.

This idea sits underneath much of modern AI. It is one reason the same base model can power many different applications once the surrounding task design and data are adapted carefully.

Follow the idea

Fine-tuning updates some or all pretrained parameters.

Self-supervised learning can supply a pretrained representation.

Related concepts (6)
Related Yenra articles (10)

Source · PyTorch tutorial on transfer learning for images

Return to Transfer Learning in the A–Z glossary