On Integrating the Data-Science and Machine-Learning Pipelines for Responsible AI

Herein, we advocate for the integration of the pipelines for data science (e.g., extraction, cleaning, and exploration) and machine learning (e.g., training data collection, feature selection, model selection, and parameter tuning), toward responsible and trustworthy artificial intelligence. We argue that the metadata generated by the machine-learning pipeline, which includes model outputs and model accuracy scores, is best managed and analyzed using data-science tools, thereby obtaining actionable insights into model performance, interpretability, and bias. We illustrate via two examples from our recent work as proof of concept: data summarization for model performance diagnostics; and input and output exploration to understand retrieval-augmented language models.