Dimension Reduction via Sum-of-Squares and Improved Clustering Algorithms for Non-Spherical Mixtures
Researchers have developed a novel method for clustering non-spherical Gaussian mixture models by employing a sum-of-squares subroutine to identify a low-dimensional projection of the data that preserves separation. This approach yields algorithms capable of clustering mixtures of centered Gaussians with significantly fewer samples and less time than previous state-of-the-art methods. The work also addresses clustering mixtures with identical, unknown covariance, and can tolerate a fraction of outliers, potentially circumventing existing lower bounds for such problems. AI
IMPACT Introduces a more efficient method for clustering complex data, potentially improving downstream AI applications that rely on data segmentation.