Towards understanding and simplifying human-in-the-loop machine learning

Machine learning application developers and data scientists spend inordinate amount of time iterating on machine learning (ML) workflows, by modifying the data pre-processing, model training, and post-processing steps, via trial-and-error to achieve the desired model performance. As a result, developers are "in-the-loop" of the development cycle. Under this "human-in-the-loop" setting, the ultimate goal of a ML system becomes shortening the time to obtain deployable models from scratch. However, some of the existing ML systems ignore this iterative aspect, and only optimize the one-shot execution of the workflow, while some of them don't provide enough support for system users to make iterative changes. Here, we first conduct a mini-survey of the applied machine learning literature to quantitatively study the user behavior in iterative ML application development. Then, we propose Helix, a declarative machine learning system implemented in Scala. Helix mainly focuses on the optimization of the execution across iterations by reusing or recomputing intermediate results as appropriate. Finally, we describe our collaboration system on top of Helix, that includes a workflow management module and a visualization tool, to make the machine learning system easier to use. In our evaluations, Helix achieved a 60% magnitude reduction in cumulative running time compared to state-of-the-art machine learning tools