Explaining Generalization Power of a DNN Using Interactive Concepts

Authors: Huilin Zhou, Hao Zhang, Huiqi Deng, Dongrui Liu, Wen Shen, Shih-Han Chan, Quanshi Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Specifically, we trained a seven-layer MLP (MLP-7-census) on the census dataset (Asuncion and Newman 2007) and a seven-layer MLP (MLP-7-TV) on the TV news dataset (Asuncion and Newman 2007), respectively. We also trained Alex Net (Krizhevsky, Sutskever, and Hinton 2017), Res Net-20 (He et al. 2016), and VGG-11 (Simonyan and Zisserman 2014) on the MNIST dataset (Le Cun et al. 1998) (Alex Net-MNIST, Res Net-20-MNIST, VGG-11-MNIST) and the CIFAR-10 dataset (Krizhevsky, Hinton et al. 2009) (Alex Net-CIFAR10, Res Net-20-CIFAR-10, VGG-11-CIFAR-10). ... Various experiments have verified our findings.
Researcher Affiliation Academia Huilin Zhou1, Hao Zhang1, Huiqi Deng1, Dongrui Liu1, Wen Shen1, Shih-Han Chan1, 2, Quanshi Zhang1* 1Shanghai Jiao Tong University 2University of California San Diego {zhouhuilin116, 1603023-zh, denghq7, drliu96, wen shen}@sjtu.edu.cn s2chan@ucsd.edu, zqs1022@sjtu.edu.cn
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The code will be released when the paper is accepted.
Open Datasets Yes Specifically, we trained a seven-layer MLP (MLP-7-census) on the census dataset (Asuncion and Newman 2007) and a seven-layer MLP (MLP-7-TV) on the TV news dataset (Asuncion and Newman 2007), respectively. We also trained Alex Net (Krizhevsky, Sutskever, and Hinton 2017), Res Net-20 (He et al. 2016), and VGG-11 (Simonyan and Zisserman 2014) on the MNIST dataset (Le Cun et al. 1998) (Alex Net-MNIST, Res Net-20-MNIST, VGG-11-MNIST) and the CIFAR-10 dataset (Krizhevsky, Hinton et al. 2009) (Alex Net-CIFAR10, Res Net-20-CIFAR-10, VGG-11-CIFAR-10).
Dataset Splits No The paper mentions 'training samples' and 'testing samples' but does not provide explicit details about train/validation/test dataset splits (e.g., percentages, sample counts, or specific split files) needed for reproduction.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details, such as library names with version numbers, needed to replicate the experiment.
Experiment Setup Yes Specifically, we trained a seven-layer MLP (MLP-7-census) on the census dataset... Each layer of the MLPs contained 100 neurons. ... We trained Alex Net, Res Net20 and VGG-11 with ρ = 0, 0.05, 0.1, 0.2, 0.3 noise on the MNIST dataset and the CIFAR-10 dataset. ... We trained a five-layer MLP on the dataset X = {0, 1}n... Here, n = 10.