#small-dataset

Question 1: Single choice

You have 300 labelled photos of everyday objects, similar to ImageNet, and an ImageNet-pretrained CNN. Which strategy is usually the safest first choice?

Fine-tune all layers with a large learning rate so they adapt quickly1
Keep the old 1,000-class output layer and map its predictions to your classes by hand2
Freeze the pretrained layers and train only a new classifier on top of their features3
Train the same architecture from scratch, since pretrained weights encode other classes4

Timed test