Transfer learning is a machine learning technique where a model trained on one task is reused, in whole or in part, as the starting point for a different but related task, rather than training a brand new model completely from scratch.
Why Start From Scratch Is Wasteful
Training a large neural network from nothing requires enormous amounts of labelled data and computing power. Many practical applications, however, do not have millions of labelled examples available. Transfer learning solves this by starting with a model that already learned general patterns from a large, related dataset, and then fine-tuning it on the smaller, specific dataset for the actual task at hand.
A Practical Example
A model originally trained to recognise thousands of general object categories in photographs has already learned to detect edges, shapes, and textures in its early layers. That same foundation can be fine-tuned with a much smaller set of medical images to recognise a specific type of scan abnormality, since the early layers do not need to be relearned from zero.
Where It Is Used Today
Transfer learning underlies most modern applications built on top of large pretrained models, including many computer vision tools and virtually all applications built using pretrained language models, where developers fine-tune a general-purpose model for a specific domain like legal or medical text rather than building a model from the ground up.