Transfer learning has two main strategies: Feature Extraction (freezing most layers and only training the final classification head) and Fine-Tuning (initializing with pretrained weights and training all layers with a smaller learning rate).