#feature-learning

Question 1: Single choice

mediumAI-checkedConvolutional Neural Networks

An MLP is trained on images by flattening each image into a vector. When you visualize its first-layer weights as images, you do not see the small edge-like filters that the first layer of a CNN typically learns. Why?

An MLP has no built-in bias toward local, translation-equivariant features, so nothing favours edge filters1
Only the last layer of an MLP gets trained, so the first layer keeps its random initial weights2
Flattening destroys the pixel values, so the first layer only receives meaningless noise3
Edge detectors need more parameters than a single dense layer of an MLP has available4

Timed test