DenseNet
DenseNet, abréviation de Densely Connected Convolutional Networks, est un type de d'apprentissage profond primarily used for classification d'image tasks. Introduced in 2017 by Gao Huang et al., DenseNet is designed to enhance the flow of information and gradients throughout the network.
La caractéristique clé de DenseNet est son schéma de connectivité unique. Dans un réseau de neurones convolutionnels (CNN), layers are connected sequentially, meaning each layer only receives input from the previous layer. In contrast, DenseNet connects each layer to every other layer that precedes it. This means that each layer has direct access to the feature maps of all previous layers, which facilitates better feature reuse and reduces the number of parameters in the model.
The architecture consists of dense blocks, where each block contains several convolutional layers. Each layer within a block takes as input the concatenated outputs of all preceding layers. This approach significantly improves the network’s ability to learn and generalize, as it encourages the network to extract a diverse set of features at different levels of abstraction.
DenseNet aborde également le problème de la disparition du gradient, a common issue in deep networks, by ensuring that gradients can flow easily through the network during backpropagation. This leads to faster convergence during training and often results in superior performance compared to traditional CNNs.
DenseNet has been applied successfully in various fields beyond image classification, including analyse d'images médicales and video processing, showcasing its versatility and effectiveness in handling complex data.