Réseau de neurones à capsules
Une capsule réseau neuronal (CapsNet) is a type of réseau de neurones artificiels designed to improve the performance of traditional réseaux de neurones convolutifs (CNNs) in understanding spatial relationships and recognizing patterns in data, particularly in traitement d'image. Proposed by Geoffrey Hinton and his team in a paper published in 2017, CapsNets aim to overcome some limitations of CNNs, such as their inability to recognize objects that have been rotated or viewed from different angles.
The core idea behind capsule networks is the concept of ‘capsules,’ which are small groups of neurons that work together to identify specific features of an object, including its orientation, position, and deformation. Each capsule outputs a vector, where the length of the vector represents the probability of the feature being present, and the orientation encodes the properties of the feature. By stacking these capsules in layers, CapsNets can capture more complex hierarchical relationships between features.
One of the key innovations of capsule networks is the dynamic routing algorithm, which helps capsules in one layer communicate with capsules in the next layer. This routing mechanism allows the network to focus on the most relevant features for recognizing an object, leading to improved accuracy in tasks like classification d'image, object detection, and segmentation.
Capsule neural networks are especially promising for tasks where the spatial relationships between features are crucial, such as in vision par ordinateur applications. They show potential in reducing the amount of training data needed and enhancing the robustness of models against variations in input data.