#domain-similarity

Question 1: Single choice

You have 500 labeled microscope images of cells, which look nothing like ImageNet photos, and a CNN pretrained on ImageNet. Following the usual rules of thumb for transfer learning, what should you try first?

Train the same architecture from scratch, since pretrained features cannot help on such data1
Freeze the network and train a linear classifier on the features just before its old output layer2
Freeze the network and train a linear classifier on activations from an earlier layer3
Fine-tune every layer end to end, since only full fine-tuning can adapt to such different images4

Timed test