Self-Supervised Generative Models

$k$-means clustering

In the previous learning problems, we had training examples with feature vectors $\mathbf{x}$ and labels $\mathbf{y}$. We now discuss an example of unsupervised learning in which no labels are given. Such data sets are very common. For example, it’s easy to take pictures with a digital camera and get huge numbers of pictures. What usually takes time is labeling the pictures, as in the ImageNet data set, or even segmenting particular objects in the picture. Thus, we can’t use the supervised training methods we have discussed so far.

However, samples of unlabeled collections still have interesting information embedded in them, and self-supervised learning has important applications. In particular, we can glean some structure from the data, which can help in representational learning and speed up supervised learning. Self-supervised or unsupervised doesn’t mean that the learning is not guided at all; the learning follows specific principles that are used to guide the organization of the data itself.