Reproducibility Study of “Studying How to Efficiently and Effectively Guide Models with Explanations”
Abstract: This paper investigates the reproducibility of the work presented by Rao et al. (2023) that explores how to guide models efficient and effectively using bounding boxes. Model guidance, achieved by jointly optimizing models with a classification and localization loss, ensures that models are right for the right reasons. Our findings indicate that the results from the original paper are reproducible. However, we observe that while these guided models attend better to the relevant objects, they do so by trading off classification performance. Furthermore, we find that some classes are attended to better than others.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Elliot_Meyerson1
Submission Number: 2207
Loading