Deep Clustering for Unsupervised Learning of Visual Features

paper

Unsupervised learning of features

Coates, and Ng uses k-means to pre-train conv nets.
Some works [21, 22, 33, 34] in paper jointly learn convnet features and image clusters with different clustering losses.

Self-supervised learning

Use pretext tasks to replace the labels annotated by humans by "pseudo-labels" directly computed from the raw input data.

Generative models

the discriminator of a GAN can produce visual features, but their performance are relatively disappointing (paper).

Approach Overview

image.png

Clustering-then-classification pattern.