In `Learning Feature Representations with K-means'  the authors apply unsupervised learning, i.e., k-means to learn 256 features (centroids, centers).
They also claim that the huge number of centers, in this case, k=256 is a limitation to this approach.
``Indeed, it is frequently best to set k as large as compute resources will
allow, considering data requirements. Though performance typically asymptotes
as k becomes extremely large, increasing the size of k is a very effective way
to squeeze out a bit of extra performance from an already-built system''