Thinking Beyond Distributions in Testing Machine Learned Models