Input Gradient
Input gradients, the rate of change of a model's output with respect to its input, are central to understanding and improving deep learning models. Current research focuses on leveraging input gradients for enhancing model robustness against adversarial attacks, generating more faithful and reliable explanations of model decisions (e.g., using Grad-CAM and its variants), and improving the interpretability of complex models like vision transformers and large language models. These efforts are significant because they address critical challenges in deploying deep learning models responsibly and reliably in various applications, from medical diagnosis to ensuring the safety of AI systems.
15papers
Papers
September 30, 2024
June 3, 2024
February 25, 2024
March 28, 2023
January 28, 2023
November 14, 2022
June 14, 2022