Interpretable Graph Capsule Networks for Object Recognition

Authors: Jindong Gu1469-1477

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical experiments show that Gra Caps Nets achieve better performance with fewer parameters and also learn disentangled representations. Gra Caps Nets are also shown to be more robust to the primary white adversarial attacks than CNNs and various Caps Nets. 4 Experiments Many new versions of Caps Nets have been proposed, and they report competitive classification performance. However, the advantages of Caps Nets over CNNs are not only in performance but also in other properties, e.g., disentangled representations, adversarial robustness. Additionally, instead of pure convolutional layers, Res Net backbones(He et al. 2016) are often applied to extract primary capsules to achieve better performance. Hence, in this work, we comprehensively evaluate our Gra Caps Nets from the four following aspects.
Researcher Affiliation Collaboration Jindong Gu University of Munich Siemens AG, Corporate Technology jindong.gu@outlook.com
Pseudocode Yes Algorithm 1: Capsule Networks; Algorithm 2: Graph Capsule Networks
Open Source Code No The paper does not contain an explicit statement about releasing the source code for the described methodology, nor does it provide a link to a code repository.
Open Datasets Yes The datasets, MNIST (Le Cun et al. 1998), F-MNIST (Xiao, Rasul, and Vollgraf 2017) and CIFAR10 (Krizhevsky et al. 2009), are used in this experiment.
Dataset Splits No The paper mentions using specific datasets and describes some aspects of the training procedure (e.g., 'set identically to (Sabour, Frosst, and Hinton 2017) (See Supplement A)'), but it does not explicitly state the percentages or sample counts for training, validation, and test splits within the main text.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments (e.g., specific GPU or CPU models).
Software Dependencies No In this experiment, the settings of these methods follow Captum package (Kokhlikyan et al. 2019) (See Supplement B). Their hyperparameter settings are default in Adversarial Robustness 360 Toolbox (Nicolae et al. 2018) (See Supplement E). While specific software packages are mentioned, their version numbers are not provided.
Experiment Setup Yes The data preprocessing, the arhictectures and the training procedure are set identically to (Sabour, Frosst, and Hinton 2017) (See Supplement A). Correspondingly, in Gra Caps Nets, 32 heads and 8D primary capsules are used. 3x3 kernels are used in Conv layers to obtain graphs with 144 nodes on MNIST, 196 nodes on CIFAR10.