Model Auditing: Faithfulness of Post-hoc Explanations Investigated

A recent study published on arXiv cs.CV has shed light on the faithfulness of post-hoc explanations for a pedestrian detection model across different domains. According to the study, researchers investigated the faithfulness of post-hoc explanations for a pedestrian detection model across different domains, employing various auditing methods, including ROI-based D-Deletion and frozen confidence terciles. The study found that deletion-based faithfulness is strongly linked to detection strength at explanation time, with a correlation coefficient of 0.70 and 0.82, respectively, for two specific models examined. This link was found to render naive confidence-stratified comparisons unreliable. Furthermore, the study revealed that the faithfulness of D-RISE was found to be domain-dependent in the central frequency range, with differences observed in frequency ranges such as f0, PIE, and JAAD. In addition, the study demonstrated that a non-perturbative EigenCAM baseline was less faithful than D-RISE but still exhibited score coupling. This research highlights the importance of auditing and verifying the faithfulness of post-hoc explanations for pedestrian detection models, given the need for explainability in intelligent vehicles to ensure safe and reliable operation. ## Auditing Methods Employed The study employed various auditing methods, including ROI-based D-Deletion and frozen confidence terciles, to investigate the faithfulness of post-hoc explanations for a pedestrian detection model. The use of these methods allowed the researchers to assess the performance of the model and identify potential biases. ## Faithfulness of D-RISE The study found that the faithfulness of D-RISE was found to be domain-dependent in the central frequency range, with differences observed in frequency ranges such as f0, PIE, and JAAD. This finding highlights the importance of considering domain-specific factors when evaluating the faithfulness of post-hoc explanations. ## Limitations and Future Work While the study provides valuable insights into the faithfulness of post-hoc explanations for a pedestrian detection model, there are several limitations that should be considered. For example, the study only examined a limited number of models and domains, and future work should aim to expand the scope of the study to include a broader range of models and domains. Additionally, the study did not explore the potential impact of biases in the training data on the faithfulness of post-hoc explanations, which should be an area of future research. ## Conclusion In conclusion, the study published on arXiv cs.CV provides valuable insights into the faithfulness of post-hoc explanations for a pedestrian detection model across different domains. The findings highlight the importance of auditing and verifying the faithfulness of post-hoc explanations for pedestrian detection models, given the need for explainability in intelligent vehicles to ensure safe and reliable operation.