What is transfer learning?
Transfer learning is the practice of taking a model trained on one task and reusing what it learned as the foundation for a different but related task. A network trained to recognize a thousand everyday object categories has already learned to detect edges, textures, and shapes. Those skills transfer to recognizing weld defects or reading meter dials, so the new model starts from knowledge instead of from random numbers.
In the most common form, a team downloads a pre-trained model, keeps most of its layers unchanged, and trains only the final layers on its own labeled data. When more data is available, the whole network can be trained further at a low learning rate, which is fine-tuning. Large language models are transfer learning at scale: pre-training on general text produces a base model, and every downstream use, from a chat assistant to a contract summarizer, builds on it.
Why transfer learning matters
Training a capable vision or language model from scratch takes millions of labeled examples and weeks of time on graphics processing units (GPUs). Most businesses have neither. Transfer learning turns a problem that needs a research lab into one that needs a few hundred to a few thousand labeled examples and a single GPU for an afternoon. It is the main reason custom models are now within reach of a mid-sized company.
Transfer learning in practice
A machine shop wants to catch surface defects on machined parts before shipment. It has about six hundred photos of defective parts and several thousand of good ones. Starting from a convolutional neural network pre-trained on ImageNet, the team replaces the final classification layer, trains on its photos, and reaches usable accuracy in a day. The same approach from scratch would have failed for lack of data.
What to watch for
- Transfer works best when the source and target data are similar; a model trained on photographs transfers poorly to X-ray images or spectrograms.
- Check the license of the source model, since some permit research use only.
- Pre-trained models carry the biases of their original training data into the new task.
- Keep a baseline model trained without transfer, so the gain can be measured rather than assumed.