Integrated Decision Gradients: Compute Your Attributions Where the Model Makes Its Decision
Authors: Chase Walker, Sumit Jha, Kenny Chen, Rickard Ewetz
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In the evaluation on Image Net, it is demonstrated that IDG outperforms IG, Left-IG, Guided IG, and adversarial gradient integration both qualitatively and quantitatively using standard insertion and deletion metrics across three common models. |
| Researcher Affiliation | Collaboration | 1Department of Electrical and Computer Engineering, University of Central Florida, Orlando, FL, USA 2Knights Foundation School of Computing and Information Sciences, Florida International University, Miami, FL, USA 3Lockheed Martin, Orlando, FL, USA |
| Pseudocode | Yes | Algorithm 1: Computing IDG with Adaptive Sampling |
| Open Source Code | Yes | Both our code and extended technical report including supplementary materials are publicly available via https://github.com/chasewalker26 /Integrated-Decision-Gradients. |
| Open Datasets | Yes | We perform our experiments in Py Torch using the 2012 validation set of Image Net (Russakovsky et al. 2015) |
| Dataset Splits | No | The paper uses the ImageNet validation set for evaluation on pre-trained models. It does not provide explicit training/validation/test splits that they created or used for training their own models within the scope of this paper. |
| Hardware Specification | Yes | We perform our experiments in Py Torch using the 2012 validation set of Image Net (Russakovsky et al. 2015) on NVIDIA A40 GPUs. |
| Software Dependencies | No | The paper mentions using PyTorch and Captum, but does not provide specific version numbers for these software dependencies. It refers to repositories for other methods but not with version specifications for their implementation. |
| Experiment Setup | Yes | Inputs are reshaped to (224, 224) for all three presented models. ... The IG and LIG attribution methods use 50 steps and a black baseline image. GIG uses the default parameters found at (Kapishnikov et al. 2021a). AGI uses the default parameters found at (Pan, Li, and Zhu 2021a). Lastly, IDG is used with 50 steps and a black baseline image. For all the methods, we use a single baseline only. |