The fundamental problem of inference in astronomy is how to reliably measure parameters from limited observational data. Recently, machine learning approaches have shown improvement upon traditional analytical methods by learning to model high order features directly from cosmological simulations. However, these data-driven methods are highly conditional on assumptions used to generate training data and their naive application could result in predictive biases when generalizing to real observations. In this talk, I will discuss the nature of machine learning solutions to observational inference problems and how we can build robust validation pipelines to ensure the reliability of our predictions. These topics will be motivated by specific examples of our work in deriving state-of-the-art dynamical mass estimates for galaxy clusters as well as our adventures in extending these methods to real observational data.