Geoffrey Hinton, the British-Canadian computer scientist who shared the 2024 Nobel Prize in Physics with John Hopfield for foundational neural network discoveries, introduced capsules in a 2011 paper on transforming autoencoders. These small groups of neurons output vectors where length indicates entity probability and direction encodes pose attributes like position and orientation. In October 2017, Hinton co-authored "Dynamic Routing Between Capsules" with Sara Sabour and Nicholas Frosst from Google Brain Toronto, presenting the architecture at NIPS. This work replaced max-pooling with "routing-by-agreement," a mechanism that allows lower-level capsules to send outputs to higher-level ones only when predictions align sharply. The approach aims to solve the "Picasso problem," where traditional convolutional networks accept distorted objects because they discard spatial relationships during feature extraction.
From Theory to Benchmark Performance
The original CapsNet architecture is shallow, using a convolution layer with 256 kernels of size 9×9 feeding into PrimaryCaps. This layer generates 1,152 eight-dimensional capsules from a 28×28 digit image, which then route to DigitCaps containing one 16-dimensional capsule per class. Training combines a margin loss function that pushes capsule lengths above 0.9 for present digits and below 0.1 for others, alongside an autoencoder-style decoder scaled by 0.0005. With 8.2 million parameters compared to 35.4 million for a baseline CNN, CapsNet achieved 0.25% test error on MNIST, outperforming the baseline’s 0.39%. On MultiMNIST, where bounding boxes overlap by 80%, it scored 5.2% against 8.1% for the CNN. You might expect these gains to translate directly to complex scenes, but on CIFAR-10, an ensemble of seven CapsNets reached only 10.6% error, matching early CNN results. The authors traced this gap to capsules struggling with cluttered backgrounds, revealing that the architecture excels at precise part-whole relationships but falters in noisy environments.
Mathematical Rigor and Enterprise Reality
Subsequent research focused on proving the convergence of dynamic routing algorithms, which minimize a concave objective function under linear constraints. A 2025 arXiv paper demonstrated that this process acts as a nonlinear gradient method, allowing for strict mathematical proof of convergence for iterative routing procedures. However, computational costs remain a significant barrier; routing runs multiple iterations over every child-parent pair, causing memory usage to grow rapidly with image size. Consequently, capsule networks never matched CNNs or vision transformers on large benchmarks like ImageNet. In 2018, Hinton and colleagues introduced matrix capsules with EM routing, replacing pose vectors with 4×4 matrices. This variant cut test errors by 45% on the smallNORB benchmark and resisted white-box adversarial attacks better than baseline CNNs. Despite these technical improvements, enterprise adoption has remained limited. Apple released code for "inverted dot-product attention" routing in 2020, which improved performance on CIFAR-10 and CIFAR-100 while using four times fewer parameters than ResNet-18. Yet, the field lacks a unified standard, and most industrial applications still rely on established convolutional architectures. Whether the current push toward unsupervised learning will finally make capsules viable for general artificial intelligence remains an open question.
