Archetypal Images in Large Photo Collections

This paper presents an approach to large scale archetypal analysis. In archetypal analysis, multivariate data points are represented as sparse convex combinations of extremal elements of a data set. It therefore allows for describing data in terms of fractions of intuitively understandable elementary concepts. However, as its computation costs grow quadratically with the number of data points, the original algorithm hardly applies to practical data analysis problems. In this paper, we present a way of considerably accelerating archetypal analysis and then apply it to search for latent structures in a large collection of images.