Transfer Learning

Question 1

What is transfer learning?

Answer 1

Starting from a pre-trained model and adapting its learned features to a new task.

Question 2

Why reuse early CNN layers in transfer learning?

Answer 2

They capture generic visual features (edges, textures) that generalize across different datasets.

Question 3

How do you adapt a pre-trained CNN for new classes?

Answer 3

Replace the original classifier head with a new dense layer sized to your target labels.

Question 4

When should you use freeze-only-head, train-all, or hybrid fine-tuning?

Answer 4

Freeze head only: small datasets for speed and less overfitting; train all: large or different-domain datasets; hybrid: start with head then unfreeze deeper layers for a balance.

Question 5

Why does transfer learning often outperform training from scratch?

Answer 5

Pre-trained features provide a significant head start, achieving high accuracy in fewer epochs on modest datasets.