Evaluating Harmony: Neural Network Explanation Metrics and Human Perception

Recent strides in deep neural network performance have catalyzed the development of state-of-the-art approaches across diverse domains. However, the inherent black-box nature of neural networks poses challenges in contexts where model explainability and transparency are paramount. In response, researchers have proposed algorithms over the years to enhance neural network understanding, offering additional insights for human experts.The surge in attribution method research underscores the need to rigorously assess and evaluate their performance. Although numerous metrics have emerged, each targeting specific properties of attribution methods, a notable gap persists in researching the alignment between these metrics and human judgment. Despite the continuous competition among researchers to produce attribution methods achieving new state-of-the-art results on established metrics, it remains unclear whether these advancements positively, negatively, or negligibly impact human comprehension of the model.This paper introduces a novel approach to evaluating the alignment between human judgment and evaluation metrics. The proposed method systematically assesses several well-established metrics across the faithfulness, robustness, and localization categories. Through this evaluation, the paper aims to shed light on the effectiveness of existing metrics in capturing human-perceived model understanding, thereby providing valuable perspectives for further advancements in the field. Our code and data are available at: https://github. com/davor10105/harmony_app