What is transfer learning?

Transfer learning is the practice of starting with a model that has already learned useful patterns from one task or dataset and adapting it to a new, related task.

Instead of training a model completely from scratch, you reuse previously learned parameters.

Pretrained model
      ↓
Reuse learned features
      ↓
Adapt to new task

Why is transfer learning useful?

Training large deep-learning models from scratch can require a lot of data and computation. A pretrained model may already contain useful representations, such as visual features in an image model.

This can help when:

• Your new dataset is relatively small.
• Training from scratch would be expensive.
• The source and target tasks share useful patterns.

Common approaches

• Feature extraction: keep most pretrained layers fixed and train only a new output layer or small task-specific head.
• Fine-tuning: start from pretrained weights and continue training some or all of the model on the new dataset.

For example, an image model pretrained on a large image dataset can be adapted to classify plant diseases using a much smaller domain-specific dataset.

Important limitation

Transfer learning works best when the pretrained model has learned representations that are relevant to the new task. If the source and target domains are very different, the benefit may be smaller.

In short, transfer learning reuses knowledge from a pretrained model to learn a new task faster and often with less data than training from scratch.