oASIS: Adaptive Column Sampling for Kernel Matrix Approximation

Computing with large kernel or similarity matrices is essential to many state-ofthe-art machine learning techniques in classification, clustering, and dimensionality reduction. The cost of forming and factoring these kernel matrices can become intractable for large datasets. We introduce an an adaptive column sampling technique called Accelerated Sequential Incoherence Selection (oASIS) that samples columns without computing the entire kernel matrix. Numerical experiments demonstrate that oASIS has performance comparable to state-of-the-art adaptive sampling methods at a fraction of the cost.