Subspace Capsule Network
Marzieh Edraki, Nazanin Rahnavard and Mubarak Shah. “Subspace Capsule Network.” 34th Conference on Association for the Advancement of Artificial Intelligence (AAAI 2020), New York, USA.
In this paper we propose the general subspace capsule network that can be successfully applied in generative as well as discriminative tasks. Subspace capsule network models the possible variations in the appearance of entities trough a group of learned capsule subspaces. Then the capsules are created by projecting the input feature vector onto these learned capsule subspaces using leaned transformations.
We evaluated the effectiveness and generalizability of SCN through a comprehensive set of experiments in three applications namely, high resolution image generation, semi-supervised image classification using the GAN framework and supervised image classification. We demonstrate the scalability of SCN to large datasets and model architectures by applying it on supervised classification of ImageNet dataset.
- High Resolution Image Generation
- Semi-supervised Image Classification
- Supervised Image Classification
For high resolution image generation, the generator of the GAN framework has been replaced by SCN, which leads to a significant improvement on the quality of the generated samples. Figure 1 shows high resolution images generated by SCN. (Figure 1: Generated samples for LSUN horse and cat with resolution 256)
The FID metric has been used to compare the quality of SCN samples of 4 datasets with prior art. (Table 1: Comparison of FID score of SCN with our baseline model. Entries with ∗ are our rerun of the baseline)
In this set of experiments the discriminator of the GAN framework has been updated with SCN network. (Table 2: Classification errors on CIFAR-10 and SVHN datasets compared with the state-of-the-art methods)
To show the effectiveness and scalability of the SCN, we also applied it on the classification task ImageNet dataset with Resnet backbone. (Table 3: Single crop, Top1 and Top 5 error rate of ImageNet classification with Resnet backbone model)