Scaling up semi-supervised learning to gigantic image collections
 Rob Fergus (CIMS)

Abstract:

With the advent of the Internet it is now possible to collect hundreds of millions of images. These images come with varying degree of label information- "clean labels" can be manually obtained on a small fraction, "noisy labels" may be extracted automatically from surrounding
text, while for most images there are no labels at all. Semi-supervised learning is a prinicipled framework for combining these different label sources. But semi-supervised learning scales polynomially with thenumber of images, which makes it impractical for hundreds of millions of images with thousands of labels.

In this paper we show how to utilize recent results in machine learning to obtain highly efficient approximations to semi-supervised learning. Specifically, we use the convergence of the eigenvectors of the normalized graph Laplacian to eigenfunctions of weighted Laplace-Beltrami operators. Our algorithm enables us to apply semi-supervised learning to clean up a huge databases of millions of images.

Joint work with: Yair Weiss (Hebrew U.) and Antonio Torralba (MIT)