Marco Tulio Ribeiro

Marco Tulio Ribeiro, machine-learning and natural-language-processing researcher. A doctoral student at the University of Washington when LIME was published at KDD 2016, later a senior researcher at Microsoft Research. His subsequent work stayed on model behaviour and testing — the Anchors method and CheckList, a behavioural-testing approach for NLP models.

Stake§

Professional and reputational. LIME launched a research agenda and his career, and the work had an interest in establishing that black boxes can be explained post hoc and trusted — the opposite bet from Rudin's.

Ribeiro is the lead author of LIME, a model-agnostic explanation method, for this topic. LIME treats every classifier as a black box and fits a simple, transparent surrogate in the local neighbourhood of a single prediction, so each output comes with an account of which features drove it. The method's promise of a "faithful" explanation is the property Lipton frames as underspecified and that Turpin and colleagues later show can fail for the step-by-step reasoning of language models.

Works in this corpus§

"Why Should I Trust You?": Explaining the Predictions of Any Classifier

¶ excerpts

Despite widespread adoption, machine learning models remain mostly black boxes. Understanding the reasons behind predictions is, however, quite important in assessing trust, which is fundamental if one plans to take action based on a prediction, or when choosing whether to deploy a new model.

— "Why Should I Trust You?": Explaining the Predictions of Any Classifier (2016)

The framing that set the post-hoc [[concept:explainability|explanation]] agenda: the model stays a black box, but a local explanation of each prediction is offered as the basis for trust.

on Explainability

We propose LIME, a novel explanation technique that explains the predictions of any classifier in an interpretable and faithful manner, by learning an interpretable model locally around the prediction.

— "Why Should I Trust You?": Explaining the Predictions of Any Classifier (2016)

LIME fits a simple, transparent model in the neighbourhood of a single prediction to approximate the black box there. The word "faithful" is the load the method is asked to carry — and the same word names the property that [[source:turpin-2023-language-models-dont-always-say-what-they-think|Turpin]] later finds chain-of-thought explanations lack.

on Explainability, Explanation faithfulness