#autoencoders

Question 1: Single choice

An autoencoder has a single hidden layer of $k$ linear units, fewer than the number of inputs, and is trained to minimize the squared reconstruction error. How does its optimal solution relate to PCA?

Its weights span the subspace of the top $k$ principal components but need not be orthogonal1
Its weights equal the top $k$ principal components exactly, orthonormal and in the same order2
It learns a curved manifold that PCA cannot represent, so its reconstruction error is lower3
Nothing in common, since through the bottleneck it simply learns to copy its input exactly4

Timed test