Transfer Learning
Transfer Learning
A technique that reuses the knowledge an AI gained in one domain to accelerate learning in a new domain.
In Simple Terms
Transfer Learning is a technique that takes an already-trained AI model as a foundation and adapts it for a new task. For example, an AI that learned to recognize 'the shapes of objects' from large amounts of images can be reused as a starting point, then further trained to distinguish between 'types of diseases.' Compared to training from scratch, it produces a highly accurate AI with far less data and in much less time.
Behind the Name
Transfer Learning combines 'transfer' — moving something from one place to another — and 'learning.' The name reflects the idea of taking knowledge an AI has already acquired while solving one task and moving it over to help with a related but different task. It is a key concept for reducing the cost of training AI.
Take a Closer Look!
Transfer Learning is a machine learning technique that applies knowledge and experience gained in one domain to learning in a different domain.
Just as humans use the logical thinking developed through studying mathematics to help learn programming, AI can reuse similar knowledge across different problems.
In most cases, an already pre-trained model is used as a base, with new task-specific layers added on top or parts of the model retrained.
For image recognition, for example, a model that already has foundational knowledge of 'colors,' 'lines,' and 'shapes' serves as the starting point.
On top of that foundation, a small amount of new knowledge is added — such as learning to 'identify apple varieties' — and the model is trained further.
A major advantage of this approach is that it enables high-performance AI even in fields where collecting large amounts of data is difficult.
Normally, teaching an AI a new concept requires tens of thousands of images, but with transfer learning, just a few hundred images can sometimes be enough to achieve solid accuracy.
It is a widely used technique across many AI development contexts, including image recognition and natural language processing.