Sparse Multi-Prototype Classification

Proceedings of the Conference on Uncertainty in Artifical Intelligence (UAI) 2018 |

We present a new class of sparse multi-prototype classifiers, designed to combine the computational advantages of sparse predictors with the non-linear power of prototype-based classification techniques. This combination makes sparse multi-prototype models especially well-suited for resource constrained computational platforms, such as those found in IoT devices. We cast our supervised learning problem as a convex-concave saddle point problem and design a provably-fast algorithm to solve it. We complement our theoretical analysis with an empirical study that demonstrates the power of our methodology.