A Unified Taylor Framework for Revisiting Attribution Methods

Authors: Huiqi Deng, Na Zou, Mengnan Du, Weifu Chen, Guocan Feng, Xia Hu11462-11469

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we empirically validate the Taylor reformulations, and reveal a positive correlation between the attribution performance and the number of principles followed by the attribution method via benchmarking on real-world datasets. We empirically validate the proposed Taylor reformulations by comparing the attribution results obtained by the original attribution methods and their Taylor reformulations. The experimental results on MNIST show the two attribution results are almost consistent. We also reveal a strong positive correlation between the attribution performance and the number of principles followed by the attribution method via benchmarking on MNIST and Imagenet.
Researcher Affiliation Academia 1 Sun Yat-Sen University 2 Texas A&M University
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statements about releasing source code or provide links to a code repository.
Open Datasets Yes We conduct the validation experiments on three models: i) Poly, a second-order polynomial model. ii) M-sg, a threelayer multi-layer perceptron (MLP) model with sigmoid activation, iii) C-sg, a three-layer CNN model with sigmoid activation. These models are all trained on MNIST 3 dataset. The footnote for MNIST is "3http://yann.lecun.com/exdb/mnist/". We also benchmark the six attribution methods in terms of infidelity and object localization accuracy on MNIST and Imagenet (Russakovsky et al. 2015). The ImageNet citation is "Russakovsky, O.; Deng, J.; Su, H.; Krause, J.; Satheesh, S.; Ma, S.; Huang, Z.; Karpathy, A.; Khosla, A.; Bernstein, M.; Berg, A. C.; and Fei-Fei, L. 2015. Image Net Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3): 211 252. doi:10.1007/s11263015-0816-y."
Dataset Splits Yes The average percentage change metrics are averaged on 3k validation set.
Hardware Specification No The paper mentions the models and datasets used for experiments (e.g., 'C-sg model', 'VGG16', 'MNIST', 'Imagenet'), but does not provide any specific hardware details such as CPU/GPU models or memory specifications.
Software Dependencies No The paper does not provide specific software dependencies with version numbers.
Experiment Setup Yes We conduct the validation experiments on three models: i) Poly, a second-order polynomial model. ii) M-sg, a threelayer multi-layer perceptron (MLP) model with sigmoid activation, iii) C-sg, a three-layer CNN model with sigmoid activation. Occlusion-p is implemented for patch size 2 2 and 4 4.