A Human-Grounded Evaluation Benchmark for Local Explanations of Machine Learning
This work addresses the need for better evaluation methods in interpretable machine learning, particularly for researchers and practitioners developing saliency explanations, but it is incremental as it builds on existing techniques like Grad-cam and LIME.
The paper tackles the problem of evaluating model saliency explanations in interpretable machine learning by proposing a human attention benchmark using aggregated multi-layer masks from human annotators for image and text domains, and results show that this benchmark is more effective than single-layer segmentation masks and reveals user biases in subjective ratings.
Research in interpretable machine learning proposes different computational and human subject approaches to evaluate model saliency explanations. These approaches measure different qualities of explanations to achieve diverse goals in designing interpretable machine learning systems. In this paper, we propose a human attention benchmark for image and text domains using multi-layer human attention masks aggregated from multiple human annotators. We then present an evaluation study to evaluate model saliency explanations obtained using Grad-cam and LIME techniques. We demonstrate our benchmark's utility for quantitative evaluation of model explanations by comparing it with human subjective ratings and ground-truth single-layer segmentation masks evaluations. Our study results show that our threshold agnostic evaluation method with the human attention baseline is more effective than single-layer object segmentation masks to ground truth. Our experiments also reveal user biases in the subjective rating of model saliency explanations.