Exploiting Sparsity in Unsupervised Classification
1We talk about two unsupervised classification projects in which the provided data have a low dimensional structure. We show how this sparsity information can be used to increase the classification accuracy. First, we describe how to use the Internet Movie Database cast-list to automatically detect and classify the actors appearing in movies such as sleepless in Seattle. We present the way we automatically detect and cluster the faces in real-time, and we show an efficient way to classify the faces in a new movie by exploiting the recent results on dictionary learning and compressed sensing. We then consider a more generic system with millions of Question and Answer pairs and use two semantic analysis methods (linked-LDA, and linked-LSA) to capture the relationship between questions and constituent words. We express improvements in performance using the recall and the F-measure. We conclude with a discussion of challenges for future work. These results were obtained at AT&T in collaboration with Howard Karloff, Patrick Haffner, Srinivas Bangalore, Taniya Mishra, and Carlos Scheidegger.