transfer learning (TL)

advancedwidespread

A training method where an AI model transfers knowledge gained from a previous task to a new, related task, so it does not need to learn from scratch.

In plain terms

Someone who knows how to ride a bicycle learns to ride a motorcycle faster: they do not need to relearn balance, cornering, or watching traffic. Transfer learning means the same for machines: the model does not start from scratch but builds on previously acquired knowledge.

This is important because training a model from scratch requires vast amounts of data, computing capacity, and money. However, if you take a large, general model as a base, a few hundred or thousand of your own examples may be enough to make it work reliably on a narrow task.

Today, almost every AI application relies on this. Large language models first learn general language patterns from vast text corpora, then apply this knowledge later to customer service responses, translation, or coding. AI agents are also built on such pre-trained models.

How it works

First, a model is trained on a large, general dataset, such as millions of images or billions of texts. This pre-training is expensive and time-consuming, but only needs to be done once.
Meanwhile, the model learns general patterns: edges, shapes, and textures in images; grammar, meaning, and context in text.
Next, the pre-trained model is adapted to the new task. Often, only the final decision-making layers are replaced or retrained, leaving the rest untouched.
The model is further trained on a smaller, targeted dataset for the new task. This is called fine-tuning and takes a fraction of the time and money of the original training.
Finally, the model is tested on separate test data to verify whether it truly performs well in the new area and has not carried over erroneous habits from the old one.

An everyday example

A plant recognition app can tell you what disease your tomato plant has from a photo of a leaf. Developers did not need to collect images of millions of diseased plants: they started with a general image recognition model and trained it further on a few thousand plant photos.

Who is affected

developerMost AI solutions today are not built from scratch but by further training a pre-trained model, saving weeks and significant costs. → Code

entrepreneur, executiveThanks to this, even a smaller company can have an AI solution tailored to its own data, not just the tech giants.

patient, doctorMedical image analysis systems often start from general image recognition models because there are few labeled images available for each specific disease. → Health

everyoneEveryday translators, chatbots, and image recognizers are almost exclusively built using this method.

A common misconception

MythTransfer learning means the model can handle any new field on its own once it has learned a lot.

In realityKnowledge transfers well only between related tasks. New areas usually require targeted adjustment (further training or at least carefully written instructions and examples), and always need verification. If the two tasks are too far apart, transferred knowledge can even worsen the result.

Limits and risks

If the original and new tasks differ too much, the model may perform worse than if it had been trained from scratch; this is called negative transfer.
Biases and errors in the base model are inherited: if the original training data was biased, the solution built on it can be too.
Due to further training on too little or one-sided proprietary data, the model may forget its previous general knowledge or overfit to the few examples.
Anyone building on another model depends on its creator: on the terms of use, the license, and exactly what data it was trained on.

Similar, but not the same

fine-tuning – Fine-tuning is a specific technique of transfer learning, while transfer learning itself is the overarching principle of applying prior knowledge to new tasks.
knowledge distillation – During distillation, a large model's knowledge is condensed into a smaller model for the same task, whereas in transfer learning, the task changes, not the model's size.
pre-training – Pre-training is the first half of the process, where general knowledge is created; transfer learning is the use of this knowledge in a new task.

Related concepts

Broader

machine learning, deep learning, model training

Narrower

fine-tuning, domain adaptation, LoRA

On the edge

pre-training, foundation model, knowledge distillation, few-shot learning, large language model

A brief history

Knowledge transfer in neural network training was already discussed in a 1976 study (Bozinovski and Fulgosi). The method became widespread in the 2010s when networks trained on large image databases began being repurposed for other tasks. In the language domain, several models in 2018 (ULMFiT, GPT), but especially BERT, made it standard practice.

Where things stand

As of late 2026, transfer learning is the default way of developing AI: most applications build on a pre-trained foundation model. When further training is needed, cheap methods that modify few parameters are typically used (for example LoRA). Large language models, however, are often not trained further at all, but adapted to the task with instructions, examples and documents attached to the query.

Status: 2026 November

In the news

Appears in 2 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Workers hide expertise when asked to train AI, fearing replacementAug 31, 2026 · Work & HR
WEF: AI is reshaping the path from junior to senior developerAug 17, 2026 · Code

Also known as

transfer learning · transfer learning · transfer learning · knowledge transfer · transfer learning · TL · knowledge transfer · pretrain–finetune

Sources

Zhuang et al.: A Comprehensive Survey on Transfer Learning (arXiv)

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.