A vast majority of localization techniques proposed for sensor networks are based on triangulation methods in Euclidean geometry. They utilize the geometrical properties of the sensor network to infer the sensor locations. A fundamentally different approach is presented in this chapter. This approach is based on machine learning, in which the authors work directly on the natural (non-Euclidean) coordinate systems provided by the sensor devices. The known locations of a few nodes in the network and the sensor readings can be exploited to construct signal-strength or hop-count based function spaces that are useful for learning unknown sensor locations, as well as other extrinsic quantities of interest. They discuss the applicability of two learning methods: the classification method and the regression method. They show that these methods are especially suitable for target tracking applications.