Convolution dilatée is a type of opération de convolution used in réseaux neuronaux, particularly in tasks involving traitement d'image and traitement du langage naturel. It modifies the traditional convolution by introducing ‘dilation’ factors, which effectively increase the size of the filter’s receptive field without adding extra parameters.
Dans la convolution standard, un filtre se déplace à travers les données d'entrée et calcule le produit scalaire at each position. This operation captures local patterns effectively. However, as the complexity of the data increases, the need for capturing wider contextual information also grows. Dilated convolution addresses this need by spacing out the filter elements, allowing it to cover a larger area of the input data.
For example, in a 1D dilated convolution with a dilation rate of 2, the filter would skip one input element between each of its weights. This means it can analyze data that is two steps apart, broadening the area of influence without increasing the number of weights in the filter. This is particularly useful for tasks like segmentation sémantique ou la synthèse audio, où la compréhension d'un contexte plus large est cruciale.
One of the significant advantages of dilated convolutions is that they can help maintain resolution in the output feature map, which is important in applications like segmentation d'image. By controlling the dilation rate, designers can fine-tune the balance between local and global feature extraction. Overall, dilated convolutions are a powerful tool in the deep learning toolkit, enabling models to learn richer representations from their input data.