Is Forward Gradient an Effective Tool for Explaining Black-box Models?
Keywords: explainability
Abstract: Gradients are widely used to explain the decisions of deep neural networks. However, as models become deeper and more complex, computing gradients becomes challenging and sometimes infeasible, hindering traditional explanation methods. Recently, the forward gradient method has garnered attention for training structure-agnostic models with discontinuous objective functions. This method perturbs only the parameters of interest for gradient computation and optimization. Inspired by this, we investigate whether the forward gradient can be employed to explain black-box models. In this work, we use the likelihood ratio method to estimate output-to-input gradients and utilize them for the explanation of model decision. Additionally, we propose block-wise computation techniques to enhance estimation accuracy. Extensive experiments in black-box settings validate the effectiveness of our method, demonstrating accurate gradient estimation and improved explainability under the black-box setting.
Primary Area: interpretability and explainable AI
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 860
Loading