An interpretable unsupervised capsule network via comprehensive contrastive learning and two-stage training

Abstract

Limited attention has been given to unsupervised capsule networks (CapsNets) with contrastive learning due to the challenge of harmoniously learning interpretable primary and high-level capsules. To address this issue, we focus on three aspects: loss function, routing algorithm, and training strategy. First, we propose a comprehensive contrastive loss to ensure consistency in learning both high-level and primary capsules across different objects. Next, we introduce an agreement-based routing mechanism for the activation of high-level capsules. Finally, we present a two-stage training strategy to resolve conflicts between multiple losses. Ablation experiments show that these methods all improve model performance. Results from linear evaluation and semi-supervised learning demonstrate that our model outperforms other CapsNets and convolutional neural networks in learning high-level capsules. Additionally, visualizing capsules provides insights into the primary capsules, which remain consistent across images and align with human vision.