STAR-Caps: Capsule Networks with Straight-Through Attentive Routing

Authors: Karim Ahmed, Lorenzo Torresani

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The experiments conducted on several image classification datasets, including MNIST, Small Norb, CIFAR-10, CIFAR-100 and Image Net show that STAR-CAPS outperforms the baseline capsule networks.
Researcher Affiliation Academia Karim Ahmed Department of Computer Science Dartmouth College karim@cs.dartmouth.edu Lorenzo Torresani Department of Computer Science Dartmouth College LT@dartmouth.edu
Pseudocode No The paper describes the architecture and mechanisms using textual descriptions and mathematical formulas, but does not provide structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any explicit statement or link regarding the availability of its source code.
Open Datasets Yes We evaluated our approach on the task of image classification using the following datasets: MNIST [15], Small Norb [16], CIFAR-10 [13], CIFAR-100 [13], and Image Net [3].
Dataset Splits No For MNIST, the dataset consists of 60K training images and 10K testing images. For CIFAR-10/100, the training set consists 50,000 images, and the testing set has 10,000 images. While 'Top-1 validation accuracy' is reported for ImageNet, the specific details for the validation split itself (e.g., size, methodology) are not provided to ensure reproducibility.
Hardware Specification No The acknowledgments section states 'We gratefully acknowledge NVIDIA and Facebook for the donation of GPUs used for portions of this work' but does not specify the models or other hardware details used for the experiments.
Software Dependencies No The paper mentions using 'Adam [12] optimizer' but does not provide specific software dependencies with version numbers (e.g., Python, TensorFlow, PyTorch versions).
Experiment Setup Yes Unless otherwise specified, STAR-CAPS models as well as EMCaps [8] models consist of a (5 5) Conv with Re LU, 1 Primary Caps, 2 Conv Caps, and 1 Class Caps. The kernel size of Conv Caps is k = 3. The number of channels of Conv and the number of capsules in each layer will be specified for each model using the following notation: #capsules={ c, n0, n1, n2, n3} as described in ( 3). We use Adam [12] optimizer, with coefficients (0.9, 0.999). The initial learning rate is 0.01, and the training batch size T = 128.