A newly proposed multi-prototype capsule network architecture outperforms existing methods in image classification accuracy across six benchmark datasets, including MNIST, C-Cube, SVHN, MCYT, CEDAR, and UTSig. The system was designed to resolve persistent training challenges in CapsNet models by guiding the network to represent complex variations in image parts during classification tasks.
Standard capsule networks operate by hierarchically identifying image parts and establishing the instantiation parameters of a complete whole, effectively performing an inverse computer graphics task. In these models, the parameters of CapsNet serve as mapping weights that transform detected parts into the whole object. However, training these networks becomes particularly challenging in scenarios that involve high intra-class or intra-part variation.
Rather than relying on a single capsule per class or part, the newly designed architecture introduces multiple co-group capsules to capture different prototypes of a given object. In the final layer of the network, these co-group capsules enter into competition, and their resulting soft outputs are used directly as training targets for a competitive cross-entropy loss function.
In the middle layers of the architecture, the most active co-group capsule is mapped to the subsequent layer using shared weights. This enables deeper capsule network structures.
Subject of Research: Multi-prototype capsule network architectures for image classification
Article Title: Deep multi-prototype capsule networks
Article References: Abbassi, S., Ghiasi-Shirazi, K., & Harati, A. (2026). Deep multi-prototype capsule networks. Multimedia Tools and Applications, 85(12), Article 860. https://doi.org/10.1007/s11042-026-22026-7
Image Credits: AI Generated
DOI: 10.1007/s11042-026-22026-7
Keywords: capsule networks, CapsNet, multi-prototype architecture, co-group capsules, image classification, competitive cross-entropy loss, shared weights, inverse computer graphics, instantiation parameters, intra-class variation, intra-part variation, MNIST, C-Cube, SVHN, MCYT, CEDAR, UTSig, deep capsule networks
News Source: Denise Maddox. (November 21, 2026). Multi-Prototype Capsule Networks Improve Image Classification Accuracy Across Multiple Benchmark Datasets. Scienmag.



