#rsa

Question 1: Single choice

A Neuromatch Academy (NMA) tutorial trains a VAE on dSprites: white shapes (heart, square or oval) with random scale, orientation and position. Its reconstructions keep these factors fairly well. But comparing the encoder's features across images shows that only images with similar positions get similar features; shape barely matters, and a linear shape classifier on these features does poorly. What does this show?

Good reconstruction does not guarantee a feature space organized by the factors a downstream task needs1
The VAE must be undertrained, since a well-trained VAE encodes every factor in its own dimension2
The decoder, not the encoder, stores the shape information, so classify the decoder weights instead3
Feature similarity cannot be measured for generative models, so the comparison is meaningless4

Timed test