Homogeneous Vector Capsules Enable Adaptive Gradient Descent in Convolutional Neural Networks
Homogeneous Vector Capsules (HVCs) are a new way to parameterize capsules that drastically reduces parameters and avoids complex routing, relying only on backpropagation weights. Using HVCs improves CNN accuracy, showing a 1.32% gain on Inception v3 compared to standard fully connected layers. This method allows researchers to use adaptive gradient descent without needing finely tuned learning rates, solving a problem in training deep networks.