1. Home
  2. Glossary
  3. Transfer learning
AI glossary · Models & training

Transfer learning

Transfer learning: Transfer learning reuses a model trained on one large task as the starting point for a different, usually smaller task, so the new model needs far less data and computing. It is the reason a general-purpose AI model can be adapted to your specific work.

Training a capable model from scratch takes enormous amounts of data and hardware. Transfer learning skips most of that. A model that has already learned general patterns, such as what edges and textures look like in photos or how English sentences are built, is taken as a base and then adjusted on a much smaller set of examples for the job you actually care about.

The classic case is vision: a network trained on millions of everyday photos gets retrained with a few hundred pictures of your product line and learns to spot defects. The same logic underlies every fine-tuned language model. A foundation model that has read a large slice of the internet is nudged with a few thousand of your support transcripts and starts answering in your company's voice, with your product names right.

Two limits to keep in mind. First, what transfers is general knowledge; the base model has never seen your data, so everything specific still has to come from your examples or your prompt. Second, whatever the base model absorbed comes along for the ride, including its biases and habits, and those can be hard to see in a small evaluation. Test the adapted model on cases that look like real work before trusting it.

Example at work

A quality engineer at a packaging plant has 400 photos of good seals and 120 of bad ones, nowhere near enough to train a vision model from nothing. Starting from a general image model and fine-tuning on those photos produces a defect detector that catches most bad seals on the line within a week of setup, and it improves as operators flag the misses.

Why it matters

Transfer learning is why small companies can use AI at all: you inherit years of expensive training and pay only for the last mile. It is also why "custom AI" claims deserve a question: what was the base, and what did you actually add?

Related terms