Few-Shot Learning by Dimensionality Reduction in Gradient Space