Réseau de capsules
Un réseau de capsules, ou CapsNet, est une architecture avancée dans le domaine de l'intelligence artificielle and apprentissage automatique, specifically in apprentissage profond. Introduced by Geoffrey Hinton and his team in 2017, Capsule Networks aim to address some of the limitations of traditional réseaux de neurones convolutifs (CNNs), en particulier dans les tâches de reconnaissance d’images.
Traditional CNNs tend to struggle with recognizing objects when they are presented in different orientations or when they are partially obscured. This is largely because they process information in a hierarchical manner, losing important spatial relationships. Capsule Networks, on the other hand, are designed to maintain these relationships through the use of capsules, which are small groups of neurons that work together to recognize specific features and their spatial orientations.
Each capsule outputs a vector whose length represents the probability of the presence of the feature it detects, while the orientation of the vector encodes the feature’s pose (e.g., position, size, rotation). This allows Capsule Networks to understand the relationship between various parts of an object, which enhances their ability to generalize from the données d'entraînement vers de nouvelles données, non vues auparavant.
One of the most notable advantages of Capsule Networks is their robustness to transformations and their ability to improve performance with fewer training samples. They are particularly useful in applications such as image recognition, where understanding the spatial hierarchy of features is crucial. Although still under research and development, Capsule Networks show promise in creating more intelligent systems that can better mimic human visual perception.