Towards credible visual model interpretation with path attribution

Authors: Naveed Akhtar, Mohammad A. A. K. Jalwana

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

Reproducibility Variable Result LLM Response
Research Type Experimental We also establish the findings empirically by evaluating the method on multiple datasets, models and evaluation metrics. Extensive experiments show a consistent quantitative and qualitative gain in the results over the baselines. and 7. Empirical Evidence
Researcher Affiliation Academia 1Computer Science and Software Engineering, The University of Western Australia, 35 Stirling highway, 6009 Crawley, Australia.
Pseudocode Yes Algorithm 1 Compute Baseline and Algorithm 2 Path integration
Open Source Code No No explicit statement about providing open-source code for the methodology described in this paper or a direct link to a repository was found. The paper mentions using 'author-provided codes for these methods' for benchmarking, but not for their own.
Open Datasets Yes Image Net (Deng et al., 2009) and CIFAR-10 (Krizhevsky et al.).
Dataset Splits Yes For each model, the results are averaged over 2,500 images from the Image Net validation set. and on 1000 images of CIFAR-10 validation set.
Hardware Specification Yes In Table 5, we report the average computational time (in seconds) required by our method and IG for both Image Net and CIFAR-10 models, computed for NVIDIA RTX 3090 with 24GB RAM using a Pytorch implementation.
Software Dependencies No The paper mentions 'Pytorch implementation' but does not provide specific version numbers for Pytorch or any other software libraries used.
Experiment Setup Yes For all the methods, we allow 150 steps. Since our technique enables the use of multiple baselines, we use 3. The reported results in the main paper, and qualitative results shown in E of this document use 150 steps, 3 baselines and δ = 5. To perform the initialization, we simply use fixed blur kernels of size 51 for Image Net images and 7 for CIFAR-10 images. We empirically noticed that with η = 1/255, the logits almost always matched reasonably well after 15 iterations.