Manifold Integrated Gradients: Riemannian Geometry for Feature Attribution

Authors: Eslam Zaher, Maciej Trzaskowski, Quan Nguyen, Fred Roosta

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments utilise deep generative models applied to several realworld image datasets. They demonstrate that IG along the geodesics conforms to the curved geometry of the Riemannian data manifold, generating more perceptually intuitive explanations and, subsequently, substantially increasing robustness to targeted attributional attacks.
Researcher Affiliation Academia 1ARC Training Centre for Information Resilience (CIRES), Brisbane, Australia 2School of Mathematics and Physics, University of Queensland, Brisbane, Australia. 3Institute for Molecular Bioscience, University of Queensland, Brisbane, Australia 4QIMR Berghofer Medical Research Institute, Brisbane, Australia.
Pseudocode Yes Algorithm 1 Geodesic Path
Open Source Code Yes The code for these experiments is available here.
Open Datasets Yes Datasets. We validate our approach on two real-image datasets: (1) the Oxford-IIIT Pet Dataset (Parkhi et al., 2012), which compromises pet images of 37 categories, making it well-suited for fine-grained classification tasks... (2) the Oxford 102 Flower Dataset (Nilsback & Zisserman, 2008), which is also used for fine-grained recognition tasks.
Dataset Splits No The paper mentions augmenting datasets during the training phase, but it does not explicitly provide percentages or sample counts for training, validation, or test splits. It refers to standard datasets, but does not specify their splits within the text.
Hardware Specification No The paper describes the VAE and classifier models used but does not provide any specific hardware details such as GPU models, CPU types, or memory specifications used for experiments.
Software Dependencies No The paper mentions using Adam optimizer but does not specify version numbers for any software libraries, frameworks, or programming languages used.
Experiment Setup Yes We train the VAEs for 150 epochs with Adam (Kingma & Ba, 2014). We then use the image reconstructions along with the labels to train classifiers based on pretrained models. All the backbones are frozen during training the classifiers for the first 10 epochs, before fine-tuning the whole models for another 7 epochs.