Transfer learning reuses a model's general knowledge to solve new problems fast.
DefinitionWhat it means
Transfer learning is the practice of pre-training a model on a broad, general dataset once, then adapting that same set of weights to a narrower downstream task with far less new data and compute. The core idea is that early layers learn general-purpose representations, such as syntax or edge detection, which transfer across tasks, while only the later layers need significant retraining. This is the mechanism that made large-scale foundation models economically viable.
Why it mattersWhy you should care
Transfer learning is why a single foundation model can be fine-tuned into a customer support bot, a coding assistant, and a medical summarizer without three separate training runs from scratch. For product teams, it collapses the cost of building a new AI feature from months of data collection to days of fine-tuning or even zero-shot prompting, which is central to how AI Foundry vendors price and package their models.
At a glanceSee it
Freeze or fine-tune turns on two questions — whether the new domain resembles the source, and whether labeled task data is plentiful.
Transfer works because early layers learn general features that reuse cleanly, while late layers stay domain-bound and usually need retraining.
Where you see itIn the wild
- Fine-tuning a foundation model on a company's support tickets
- Research papers describing pretrain-then-adapt pipelines
- Vendor docs explaining why a base model needs little task-specific data