LE-CapsNet: A Light and Enhanced Capsule Network

2025-11-12

AI Summary

To address the high computational cost, excessive parameter count, slow inference speed, and inferior accuracy of Capsule Networks (CapsNets) relative to CNNs, this paper proposes a lightweight and efficient architectural improvement. Our method streamlines the network topology, optimizes convolutional layer design and dynamic routing, and incorporates weight regularization. The resulting model reduces parameters to 3.8 million and achieves a 4× speedup in inference latency. Crucially, it preserves CapsNets’ inherent robustness to affine transformations while attaining 76.73% test accuracy on CIFAR-10 and 94.3% on AffNIST—substantially outperforming the original CapsNet. To our knowledge, this is the first work to achieve accuracy superiority over baseline CapsNets on standard benchmarks without compromising either efficiency or transformation robustness, establishing a new paradigm for deployable, lightweight capsule networks.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Natural Language Processing: Safety and Robustness
• Machine Learning: Adversarial Learning & Robustness

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods

Abstract

Capsule Network (CapsNet) classifier has several advantages over CNNs, including better detection of images containing overlapping categories and higher accuracy on transformed images. Despite the advantages, CapsNet is slow due to its different structure. In addition, CapsNet is resource-hungry, includes many parameters and lags in accuracy compared to CNNs. In this work, we propose LE-CapsNet as a light, enhanced and more accurate variant of CapsNet. Using 3.8M weights, LECapsNet obtains 76.73% accuracy on the CIFAR-10 dataset while performing inference 4x faster than CapsNet. In addition, our proposed network is more robust at detecting images with affine transformations compared to CapsNet. We achieve 94.3% accuracy on the AffNIST dataset (compared to CapsNet 90.52%).

Problem

• Improving CapsNet speed while maintaining detection accuracy
• Reducing computational resource requirements of Capsule Networks
• Enhancing robustness in detecting transformed and overlapping images

Innovation

• Lightweight CapsNet with reduced computational complexity
• Enhanced accuracy through optimized network architecture
• Faster inference speed with fewer model parameters