Quantitative Evaluation of Machine Learning Explanations: A Human-Grounded Benchmark
summary
Proposes a repeatable benchmark that aggregates selection masks from 10 annotators into gradated human attention maps and compares them to saliency outputs from Grad-CAM and LIME using ratings from 200 raters. Claims the threshold-agnostic, multi-layer baseline aligns differently than single-layer segmentation and often yields lower mean absolute error versus single-layer ground truth. Reports a floor effect in ratings, higher MAE for LIME’s chunky superpixels, and stronger user sensitivity to false-negative than false-positive explanation pixels.
AI-written summary · about
related talks
01h 05mEvaluating Model Explainability Methods
45mExplainable Machine Learning Explained
48mInherent Trade-offs with the Local Explanations Paradigm
16mBuilding Trust through Explainable AI
22mInterpreting Machine Learning Models with LIME
30mTowards a more transparent AI - Decrypting AI models using LIME
14mThe effects of example-based explanations in a machine learning interface
48mInterpretability - now what?