Training a model from randomly initialized weights means it has to discover everything from scratch, purely from whatever labelled examples are available for that specific task. Transfer learning avoids most of that work by starting from a model already trained on a large, related source task, then fine-tuning it on a much smaller labelled dataset for the actual target task.
Why pretrained features transfer
Features learned on a large source task, edge and texture detectors in early vision layers, general syntactic patterns in language layers, are often genuinely useful for a related target task, even one with far less labelled data of its own. Fine-tuning adjusts these existing features toward the new task rather than learning everything from nothing, which is the source of transfer learning's efficiency.
Fine-tuning versus feature extraction
There are two common ways to use a pretrained model. Fine-tuning continues training some or all of the pretrained weights on the new task's data. Feature extraction freezes the pretrained weights entirely and only trains a new final layer on top. Fine-tuning usually reaches higher accuracy given enough target data, while feature extraction is cheaper and works better with very little target data.
When transfer learning helps most
The benefit scales with how similar the source and target tasks actually are. Highly related tasks see dramatic reductions in the labelled data needed to reach strong accuracy. Weakly related tasks see a smaller benefit, and in rare cases a poorly matched source task can even slow learning down relative to starting fresh.
Try it yourself
The Transfer Learning Lab lets you adjust both the amount of labelled target data and a task-similarity setting, and compare fine-tuned accuracy against training from scratch as both curves converge.
🧪 Try it yourself: the Transfer Learning Lab simulation lets you experiment with everything described above directly in your browser.