Briefly describe the curse of dimensionality. Why can k-NN still perform well on some datasets like the handwritten digits you considered in the homework, despite these images being very high dimensional?