#encoder

Question 1: Single choice

You have a million unlabeled images but only 2,000 labeled ones. You first train a deep autoencoder on all the images. How should you use it to build a classifier?

Reuse its encoder as the classifier's lower layers and add a new output layer1
Reuse the decoder as the classifier's lower layers, since it was trained last2
Train the classifier on the autoencoder's reconstructions instead of the original images3
Use the reconstruction error of each unlabeled image as its class label4

Timed test