Data Science and Machine Learning applications are very different from traditional applications, and the differences make it difficult to apply traditional techniques for writing requirements and performing QA. How do you write requirements and test an application that inherently has an element of randomness built into it? How do you test a neural network when you can't explain why one set of neural weights performs better than another set? Because of this, Data Science and Machine Learning have been resistant to the application of both traditional requirements and traditional QA.