Local Path Integration for Attribution

Authors: Peiyu Yang, Naveed Akhtar, Zeyi Wen, Ajmal Mian

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental With extensive experiments on the validation set of Image Net 2012 (Russakovsky et al. 2015) using two visual classification models, we show that the proposed method of Local Path Integration (LPI) consistently outperforms the existing path-based attribution methods. We also contribute an evaluation metric for reliable performance estimation of the attribution methods.
Researcher Affiliation Academia Peiyu Yang1, Naveed Akhtar1, Zeyi Wen*2,3, Ajmal Mian1 1 The University of Western Australia 2 Hong Kong University of Science and Technology (Guangzhou) 3 Hong Kong University of Science and Technology peiyu.yang@research.uwa.edu.au, naveed.akhtar@uwa.edu.au, wenzeyi@ust.hk, ajmal.mian@uwa.edu.au
Pseudocode No The paper includes mathematical equations describing the method but does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/ypeiyu/LPI.
Open Datasets Yes We apply Diff ID to evaluate the performance of attribution methods on VGG-16 (Simonyan and Zisserman 2015) and Res Net-34 (He et al. 2016) on Image Net 2012 validation set (Russakovsky et al. 2015).
Dataset Splits Yes With extensive experiments on the validation set of Image Net 2012 (Russakovsky et al. 2015) using two visual classification models, we show that the proposed method of Local Path Integration (LPI) consistently outperforms the existing path-based attribution methods.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models or processor types used for running experiments.
Software Dependencies No The paper mentions models like VGG-16 and Res Net-34 but does not list specific versions of ancillary software or libraries (e.g., PyTorch, TensorFlow, Python version).
Experiment Setup Yes For IG, we segment the linear path with 20 steps for the integral. In AGI, the reference is generated by the PGD attack (Madry et al. 2018) with 20 steps and a single random target class. For both LPI and EG, we employ 20 references for each input with one random step. In LPI, we empirically divide the learned distributions into 9 and 7 neighborhoods for VGG-16 and Res Net-34, respectively.