What is transfer learning?
Transfer learning takes a model that already learned useful representations on a large source task and reuses it for a new target task. The intuition is that features learned on broad data, such as edges, textures, and object parts, are useful far beyond the original task, so you rarely need to learn everything again from scratch.
In practice this means starting from a pretrained model and adapting it, which is faster and needs far less labeled data than training a fresh model.
Key takeaways
- Transfer learning reuses representations from a pretrained model.
- It reduces the labeled data and compute a new task requires.
- It is most effective when source and target tasks are related.
How it works
You take a pretrained model, often a backbone trained on a large dataset, and adapt it to the target task. Options range from using its frozen features as inputs to a small new head, to fine-tuning some or all of its weights on the target data. The closer the source and target domains, the more of the pretrained knowledge transfers cleanly.
Why it matters
Transfer learning is the practical foundation of modern applied machine learning, because it turns the enormous cost of pretraining into a reusable asset. It is what lets teams reach strong performance on niche tasks with modest datasets, which is central to lowering the cost of building visual and multimodal systems.
Frequently asked questions
What is the difference between transfer learning and fine-tuning?
Transfer learning is the general idea of reusing a pretrained model. Fine-tuning is one way to do it, by continuing to train the model's weights on the new task.
When does transfer learning work best?
When the source and target tasks share structure, so the learned features remain relevant to the new problem.
Related terms