GLIME: General, Stable and Local LIME Explanation
Authors: Zeren Tan, Yang Tian, Jian Li
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments are conducted on the Image Net dataset. Specifically, we randomly choose 100 classes and select an image at random from each class. The models chosen for explanation are Res Net18 [13] and the tiny Swin-Transformer [18] (refer to Appendix A.7 for results). Our implementation is derived from the official implementation of LIME2. Metrics. (1) Stability: To gauge the stability of an explanation method, we calculate the average top-K Jaccard Index (JI) for explanations generated by 10 different random seeds. |
| Researcher Affiliation | Academia | Zeren Tan Tsinghua University thutzr1019@gmail.com Yang Tian Tsinghua University tyanyang04@gmail.com Jian Li Tsinghua University lapordge@gmail.com |
| Pseudocode | No | No structured pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | Yes | 1Code is available at https://github.com/thutzr/GLIME-General-Stable-and-Local-LIME-Explanation |
| Open Datasets | Yes | Our experiments are conducted on the Image Net dataset1. The experiments use images from the validation set of the Image Net-1k dataset. |
| Dataset Splits | Yes | The experiments use images from the validation set of the Image Net-1k dataset. To ensure consistency, a fixed random seed (2022) is employed. Specifically, 100 classes are uniformly chosen at random, and for each class, an image is randomly selected. |
| Hardware Specification | No | No specific hardware details (e.g., GPU models, CPU types, memory) used for running the experiments are mentioned in the paper. |
| Software Dependencies | No | The quickshift method from scikit-image is then employed to segment images into super-pixels... The pretrained models used are sourced from torchvision.models... While software components are mentioned, specific version numbers for these libraries (scikit-image, torchvision) are not provided, which is necessary for reproducibility. |
| Experiment Setup | Yes | The initial step involves cropping each image to dimensions of (224, 224, 3). The quickshift method from scikit-image is then employed to segment images into super-pixels, with specific parameters set as follows: kernel_size=4, max_dist=200, ratio=0.2, and random_seed=2023. ... The hide_color parameter is configured as None... The distance_metric is explicitly set to l2... The default value for alpha in Ridge regression is 1... Ten different random seeds are utilized to compute explanations for each image. The random_seed parameter in both the Lime Image Explainer and the explain_instance function is set to these specific random seeds. |