#hog

Question 1: Single choice

Before 2012, a typical image classifier computed SIFT or HOG descriptors and trained a classifier such as an SVM on them. What is the key difference in a CNN such as AlexNet?

Its features are learned from data, layer by layer, jointly with the classifier1
The CNN needs no labelled images, while the SIFT/HOG pipeline needs labels2
It feeds SIFT or HOG descriptors into a deep network, so only the classifier on top changes3
Its first layer holds fixed edge detectors such as Sobel kernels, and only deeper layers are trained4

Timed test