Fine-Tuning vs Feature Extraction
What you'll learn: You'll understand the two main strategies for using pretrained models—either as frozen feature extractors or by updating their weights through fine-tuning.
The Core Difference
When applying transfer learning with pretrained models (like ResNet or EfficientNet trained on ImageNet), you have two fundamental approaches:
Feature Extraction treats the pretrained model as a fixed feature transformer. You freeze all the convolutional layers (preventing weight updates during training) and only train new layers you add on top—typically the final classification head. The pretrained model acts like a sophisticated preprocessor that converts raw images into rich feature representations.
Fine-Tuning allows the pretrained weights to update during training. You start with the learned weights but continue training them (often with a small learning rate) on your specific dataset. This adapts the entire network—or selected layers—to your particular task.
When to Use Each Approach
Use Feature Extraction when:
- Your dataset is small (high risk of overfitting if you update millions of parameters)
- Your task is similar to the original task (ImageNet features transfer well)
- You have limited computational resources
Use Fine-Tuning when:
- You have a reasonably sized dataset (thousands+ of images)
- Your task differs significantly from ImageNet classification
- You want maximum performance and have the compute budget
Analogy
Think of a pretrained model as a professional photographer's trained eye. Feature extraction is like hiring them to take photos (frozen expertise) while you decide what to do with those photos. Fine-tuning is like working alongside them, gradually adjusting their shooting style to match your specific aesthetic preferences.
Key Takeaway: Feature extraction freezes pretrained weights and trains only new layers, while fine-tuning updates the pretrained weights themselves—choose based on dataset size, task similarity, and computational resources.