Towards Conceptual Models for Machine Learning Computations

We make the case for conceptual models that give the human designer full visibility and control over key aspects of ML applications, including input data preparation, training and inference of the ML models. Our models aim to: (i) achieve better documentation of ML analytics (ii) provide a foundation for a chain of trust in the ML analytics outcome (iii) provide a lever to enforce ethical and legal constraints within the ML pipeline. Representational models can dramatically increase reusability of large-scale ML analytics, while decreasing their roll-out time and cost. Also, they will support novel solutions to time-honored issues of analytics like non-uniform data veracity, privacy and latency profiles.