A new approach learns from conditional distributions with limited samples.
problem Learning from conditional distributions with few samples.
method Min-max reformulation and Embedding-SGD algorithm.
result Significant improvement over existing algorithms in experiments.