Freezing Layers in Transfer Learning

Medium

neural-networks NVIDIA Meta Google

You're fine-tuning a pretrained image classification network on a small, domain-specific dataset (a few hundred images). Should you freeze the early convolutional layers, the later layers, or train the whole network end-to-end? Explain the tradeoff.

Multiple Choice

Always train the entire network end-to-end regardless of dataset size, since pretrained weights are just a starting point Freeze the early layers, which learn generic transferable features, and fine-tune only the later, task-specific layers — a small dataset risks overfitting and destroying useful pretrained features if the whole network is unfrozen Freeze the later layers and fine-tune only the early layers, since early layers change the most between tasks Freezing layers has no effect on outcomes and only changes training speed

Correct!

Incorrect — try again next time!

Upgrade to Pro to unlock solutions and explanations

Freezing Layers in Transfer Learning

Multiple Choice

Related problems

Deep-dive guides