Capsule Networks are neural networks consisting of so-called capsules – groups of neurons that encode not only the probability of a feature but also its properties such as position and orientation.
Unlike classical Convolutional Neural Networks (CNNs), Capsule Networks better capture hierarchical relationships and spatial information.
Communication between capsules occurs via a dynamic routing-by-agreement method, where capsules in higher layers only receive signals from lower capsules that make similar predictions.
This leads to more robust object recognition under different perspectives and distortions.
Capsule Networks are being researched to complement or replace classical CNN models in areas such as computer vision and pattern recognition.