What is transfer learning and when is it useful?

• A. Reusing a pre-trained model on a related task to improve performance with limited data.
• B. Training from scratch with abundant data.
• C. Using random weights.
• D. Freezing all layers and not training.

Answer: (A)

Transfer learning is about taking a model that’s already learned useful representations from one task and applying it to a related, but different, task. The typical approach is to start with a pre-trained model that was trained on a large dataset and adapt it to the new problem, either by using it as a fixed feature extractor (freeze most layers and train a new output layer) or by fine-tuning parts of the network on the new task. This is especially helpful when you have limited labeled data for the new task, because the model already encodes general features that transfer to related problems. You can often get good performance with much less data than would be needed to train from scratch, and it’s common to reuse representations in domains like vision and natural language processing where large pre-trained models exist. The other options don’t fit transfer learning: training from scratch uses no prior knowledge; starting with random weights offers no learned features to transfer; freezing all layers and not training means the model won’t adapt to the new task.